Open-weight AI model directory
412 models you can download and run yourself, with the license that covers the weights and whether commercial use is allowed. Every row comes from a sourced AI Wiki article.
- 222 permissive
- 105 custom license
- 51 non-commercial
- 34 partly open
Open-weight models
412 models
ERNIE 5.5 Flash Base
Custom licenseBaidu · Language · Nov 5, 2026
- License
- ERNIE 5.5 Flash Base Model License Agreement v1.0
- Commercial
- Allowed with conditions
- Parameters
- 280B total, 14B active (about 156 GB at 4-bit)
Excludes the EU, UK and South Korea; commercial use needs notice and attribution up to US$50M revenue and written authorization above it.
Inkling 2
PermissiveThinking Machines Lab · Language · Nov 5, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- Inkling 2 Small: 594B total, 17B active per token (about 332 GB at 4-bit)
Inkling 2 Small weights on Hugging Face since November 12, 2026; hosted-only on Tinker for the first week after launch.
Weights on Hugging FaceMemory estimateArticle fact-checked
Hydrogen-1.0
PermissiveDeep Principle · Science · Nov 2, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- About 753B (753,329,940,480 in the BF16 checkpoint)
The Hugging Face repository is gated: users accept access conditions before download.
LongCat-2.5
PermissiveMeituan LongCat team · Language · Nov 2, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 1.6T total, about 48B active (about 894 GB at 4-bit)
The hosted launch came first; the weights were published on Hugging Face in November 2026.
Ripple-65B Preview
PermissiveRicursive Intelligence · Language · Nov 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 65B (about 36 GB at 4-bit)
Pika H3 Flex
Custom licensePika (Mellis, Inc.) · Video · Oct 29, 2026
- License
- MiniMax H3 Community License Agreement
- Commercial
- Allowed with conditions
- Parameters
- 33B base generator (MiniMax H3) with two rank-128 LoRA adapters
Excludes the EU, UK, South Korea and US; commercial use above US$20M annual revenue needs MiniMax's written authorization.
Weights on Hugging FaceArticle fact-checked
Nemotron 3.5 Transcribe
PermissiveNVIDIA · Speech and audio · Oct 28, 2026
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- About 34B total, about 4.3B active (including the audio encoder)
Weights on Hugging FaceArticle fact-checked
LFM3
Custom licenseLiquid AI · Language · Oct 27, 2026
- License
- LFM Open License v1.0
- Commercial
- Allowed with conditions
- Parameters
- 1.73B to 2.17B (six post-trained models and the LFM3-1.7B-Base checkpoint) (about 1.0 GB at 4-bit)
Commercial use is licensed only for entities under $10M annual revenue; qualified non-profits doing research are exempt from the threshold.
Memory estimateArticle fact-checked
Nanbeige
PermissiveNanbeige LLM Lab (BOSS Zhipin / Kanzhun Limited) · Language · Oct 25, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 33.1B dense (Nanbeige4.3-33B); earlier Nanbeige4 to 4.2 releases are 3B-class models (about 18 GB at 4-bit)
All Nanbeige4 and later checkpoints are Apache 2.0. Nanbeige-16B (2023) required an application for commercial use.
Weights on Hugging FaceMemory estimateArticle fact-checked
Kimi K3.1 Pro
Custom licenseMoonshot AI · Language · Oct 23, 2026
- License
- Kimi K3.1-Pro License
- Commercial
- Allowed with conditions
- Parameters
- 2.8T total, 104B active (about 1565 GB at 4-bit)
Model-as-a-service firms over $20M yearly revenue need a separate agreement; very large products must display Kimi K3.1-Pro.
Weights on Hugging FaceArticle fact-checked
Nemotron 3.5 Bolt E4B
PermissiveNVIDIA · Language · Oct 23, 2026
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- 8B total (4B effective) (about 4.5 GB at 4-bit)
Post-trained from Google's Gemma 4 E4B, whose own weights are under Apache 2.0.
Weights on Hugging FaceMemory estimateArticle fact-checked
WeVisOne-VL
PermissiveTencent (WeChat Vision team) · Vision · Oct 22, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 2B and 4B variants, each with a 9.6M pixel decoder
AW-0
PermissiveResearchers at UC Berkeley, NVIDIA and the University of Pennsylvania · Robotics · Oct 21, 2026
- License
- BSD 2-Clause
- Commercial
- Allowed
Recursive Flint
PermissiveRecursive Superintelligence · Language · Oct 21, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 39.8B total; approximately 2.5B active per token (about 22 GB at 4-bit)
Spark-X2.6-4B
PermissiveSparkLLM / XHToken · Language · Oct 21, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 4,112,079,360 (about 2.3 GB at 4-bit)
Ornith-1.5-122B
PermissiveOrnith (ornith-ai) · Language · Oct 20, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 125.1B in checkpoint metadata (named 122B) (about 70 GB at 4-bit)
MiniMax M3.1 Flash
Custom licenseMiniMax · Language · Oct 19, 2026
- License
- MiniMax Community License
- Commercial
- Allowed with conditions
- Parameters
- About 428B total, 23B active per token (about 239 GB at 4-bit)
Commercial use needs a Built with MiniMax M3.1-Flash notice and a notice to MiniMax; written authorization above $20M yearly revenue.
Weights on Hugging FaceMemory estimateArticle fact-checked
Beam (Reflection)
PermissiveReflection AI · Language · Oct 16, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 501.3B total, 23.4B active (about 280 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Muse Spark Open
Custom licenseMeta Superintelligence Labs · Language · Oct 15, 2026
- License
- Muse Model License
- Commercial
- Allowed with conditions
- Parameters
- 816B total, 42B active (about 456 GB at 4-bit)
Gated repository. Products must show Built with Muse Spark Open; API providers above $20M yearly revenue need an agreement with Meta.
Weights on Hugging FaceMemory estimateArticle fact-checked
Nemotron 3.5 Bolt
PermissiveNVIDIA · Language · Oct 15, 2026
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- 74B total, 9B active (about 41 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Nemotron 3.5 Super VL
PermissiveNVIDIA · Language · Oct 15, 2026
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- 120B total, 12B active per token (about 67 GB at 4-bit)
Qwen-Image-2.1-Turbo
Non-commercialQwen team (Alibaba) · Image · Oct 9, 2026
- License
- Qwen Research License
- Commercial
- Not allowed
- Parameters
- 7B (visual generation component)
Research and evaluation use only; commercial use needs a separate license from Qwen.
Weights on Hugging FaceArticle fact-checked
Underdog Saluki 27B
Non-commercialUnderdog (Conway Research) · Language · Oct 7, 2026
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 27B (2-bit quantized; 7.89 GB GGUF file) (about 15 GB at 4-bit)
Relicensed from Apache 2.0 to CC BY-NC 4.0 on November 5, 2026; commercial use needs a separate license from Conway Research.
Weights on Hugging FaceMemory estimateArticle fact-checked
EmbeddingGemma 2
PermissiveGoogle DeepMind · Language · Oct 6, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 740M (all modality encoders loaded) (about 0.4 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Mistral Large 4
Custom licenseMistral AI · Language · Oct 6, 2026
- License
- Mistral Ecosystem Open-Weight License
- Commercial
- Allowed with conditions
- Parameters
- About 1T total, 49B active per token (52B counting embeddings and output layers) (about 559 GB at 4-bit)
Commercial use allowed, but Model-as-a-Service operators need a separate commercial license from Mistral.
Weights on Hugging FaceMemory estimateArticle fact-checked
Decision 2.0
PermissivevLLM Semantic Router project · Other · Oct 3, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 0.60B to 29.37B (six checkpoints, nominal sizes 0.6B to 27B)
Article fact-checked
Kolibri 1
PermissiveAleph Alpha · Language · Oct 3, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 78.1B total, 3.46B active (excluding embeddings) (about 44 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Whistle (Cactus Compute)
PermissiveCactus Compute · Speech and audio · Oct 2, 2026
- License
- Apache 2.0
- Commercial
- Allowed
Weights on Hugging FaceArticle fact-checked
SenseNova 6.8 Code2Video
PermissiveSenseTime (SenseNova team) · Video · Oct 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 27B (27,356,728,560)
Weights on Hugging FaceArticle fact-checked
Intern-Decision
PermissiveInternLM team, Shanghai AI Laboratory · Vision · Sep 26, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 0.8B, 2B and 4B (about 0.85B, 2.21B and 4.54B in the published safetensors)
Three sizes, all Apache 2.0; the upstream Qwen3.5 license file (also Apache 2.0) is kept alongside.
Weights on Hugging FaceArticle fact-checked
Ming-Image-0.1-Design
PermissiveinclusionAI (Ant Group) · Image · Sep 22, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 6B diffusion transformer per model (developer figure); the checkpoint also holds a 17B multimodal language model and a 1.5B connector
Weights on Hugging FaceArticle fact-checked
Alice AI Foundation 80B-A3B
PermissiveYandex · Language · Sep 21, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 80B total, 3B active (about 45 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Xiaomi MiMo-V2.6
PermissiveXiaomi (MiMo team) · Language · Sep 21, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 310B total, 15B active (Flash); 1.02T total, 42B active (Pro) (about 173 GB at 4-bit)
Pro and Flash each ship as RL and MOPD checkpoints, all MIT. RL environments and training code were released too.
Memory estimateArticle fact-checked
Qwen-Image-2.1
Non-commercialQwen team, Alibaba · Image · Sep 20, 2026
- License
- Qwen Research License
- Commercial
- Not allowed
- Parameters
- 7B diffusion transformer; 8B Qwen3-VL text encoder
Research and evaluation use only; commercial use needs a separate license from Qwen.
Weights on Hugging FaceArticle fact-checked
Step 5 Preview
Custom licenseStepFun · Language · Sep 20, 2026
- License
- StepFun Model License 1.0
- Commercial
- Allowed with conditions
- Parameters
- 600B total, 27B active (about 335 GB at 4-bit)
Device makers and MaaS operators above US$20M yearly revenue need a separate commercial license; large services must show attribution.
Weights on Hugging FaceMemory estimateArticle fact-checked
DeepSeek V4.1-Flash
PermissiveDeepSeek · Language · Sep 10, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 552B backbone (plus 196B Engram memory parameters) (about 308 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
MiniCPM5-2B
PermissiveOpenBMB · Language · Sep 7, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 2.5B (2,516,756,480 total; 1,981,982,720 non-embedding) (about 1.4 GB at 4-bit)
The datasets named on the model card (UltraData family) are published on Hugging Face under Apache 2.0.
Weights on Hugging FaceMemory estimateArticle fact-checked
Extropic Z1T
Custom licenseExtropic · Other · Sep 4, 2026
- License
- License not stated
Weights repo Extropic-AI/Z1T-0 states no license; the code repo extropic-ai/sparse-transformers is Apache 2.0.
Weights on Hugging FaceArticle fact-checked
Ling-3.0-flash-Sante
PermissiveinclusionAI · Science · Sep 4, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 124B total, 5.1B active per token
Weights were published on Hugging Face and ModelScope after the September 2026 API-only launch. Research model, not a medical device.
Weights on Hugging FaceArticle fact-checked
Hy4 Preview
PermissiveTencent Hy Team · Language · Aug 28, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 770B total, 49B active (backbone) (about 430 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
GLM-5.3-Flash
PermissiveZ.ai · Language · Aug 26, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 320B total, 18B active (about 179 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Qwen3.8-Flash-Next
Custom licenseQwen team, Alibaba · Language · Aug 26, 2026
- License
- Qwen Community License 1.0
- Commercial
- Allowed with conditions
- Parameters
- 125B main model, 6B active; plus a 51B N-gram table and a 4B multi-token-prediction module (about 70 GB at 4-bit)
Model-as-a-service and AI work-assistant businesses need a separate Qwen license; large commercial products must display the model name.
Weights on Hugging FaceMemory estimateArticle fact-checked
WeMM-Embedding
PermissiveTencent (WeChat Vision team) · Vision · Aug 25, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 9.41B (9,407,831,280 in the published checkpoint)
Weights on Hugging FaceArticle fact-checked
GLM-5.3
Custom licenseZ.ai · Language · Aug 14, 2026
- License
- GLM-5.3 License
- Commercial
- Allowed with conditions
- Parameters
- 744B total, 40B active (Z.ai model-index label) (about 416 GB at 4-bit)
Model-as-a-service operators with over US$10 billion in 12-month revenue must pass a Z.ai security review before commercial use.
Weights on Hugging FaceMemory estimateArticle fact-checked
dots3-note Preview
Permissivedots studio (Xiaohongshu) · Language · Aug 13, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 280B total, 16B active (language backbone) (about 156 GB at 4-bit)
Nemotron 3.5 Lightning
PermissiveNVIDIA · Language · Aug 11, 2026
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- 30B total, 3B active (31.6B / 3.6B counting embeddings) (about 17 GB at 4-bit)
NVIDIA published pre-training and post-training dataset collections and recipes alongside the weights.
Weights on Hugging FaceMemory estimateArticle fact-checked
Muse Glimmer
PermissiveMeta AI · Language · Aug 10, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 29.6B (29,776,626,688 in the BF16 checkpoint) (about 17 GB at 4-bit)
Meta also publishes a separate Muse Glimmer Usage Policy that lists prohibited uses.
Weights on Hugging FaceMemory estimateArticle fact-checked
MAGI-2 Preview
PermissiveSand.ai · Video · Aug 5, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 114B total, about 6B active
The bundled audio VAE is Stable Audio Open 1.0, which upstream carries the Stable Audio Community License.
Qwen3.8-Max
Custom licenseQwen team, Alibaba · Language · Aug 3, 2026
- License
- Qwen3.8-Max License
- Commercial
- Allowed with conditions
- Parameters
- 2.4T total, 95B active (about 1341 GB at 4-bit)
Open checkpoint is text only and requires thinking mode. Model-as-a-service firms above US$50M yearly revenue need a separate Qwen license.
Weights on Hugging FaceArticle fact-checked
BigBang-v1
PermissiveThe Endless Frontier team · Language · Aug 2, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 35B total, about 3B active (about 20 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
MiniMax H3
Custom licenseMiniMax · Video · Jul 31, 2026
- License
- MiniMax H3 Community License Agreement
- Commercial
- Allowed with conditions
- Parameters
- 33B (H3-Omni-Transformer generator)
License excludes the EU, UK, South Korea and US. Over $20M yearly revenue needs authorization. Context-IR and Regenerate-2K stages are not released.
Weights on Hugging FaceArticle fact-checked
Frontis-MA1
Non-commercialFrontisAI (Horizon Research of Frontis.AI with Tsinghua University) · Other · Jul 30, 2026
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 35B total, about 3B active (Frontis-MA1-35B)
Weights on Hugging FaceArticle fact-checked
Ling-3.0-flash
PermissiveinclusionAI · Language · Jul 23, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 124B total, 5.1B active per token (about 69 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Qwen3.8
Custom licenseQwen team, Alibaba Cloud · Language · Jul 19, 2026
- License
- Qwen3.8-Max License
- Commercial
- Allowed with conditions
- Parameters
- 27.8B (Qwen3.8-27B); 2.4T total, 95B active (Qwen3.8-Max) (about 16 GB at 4-bit)
Applies to the flagship Qwen3.8-2.4T-A95B; Qwen3.8-27B is Apache 2.0. Hosted qwen3.8-max snapshots add features the open checkpoint lacks.
Memory estimateArticle fact-checked
Kimi K3
Custom licenseMoonshot AI · Language · Jul 16, 2026
- License
- Kimi K3 License
- Commercial
- Allowed with conditions
- Parameters
- 2.78T total; 104.2B activated per token (about 1553 GB at 4-bit)
Model-as-a-service firms over $20M yearly revenue need a separate agreement; very large products must display Kimi K3.
Weights on Hugging FaceArticle fact-checked
Inkling (Thinking Machines Lab)
PermissiveThinking Machines Lab · Language · Jul 15, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 975B total, 41B active (about 545 GB at 4-bit)
A separate Model Acceptable Use Policy restricts some uses. Inkling-Small is also open under Apache 2.0.
Weights on Hugging FaceMemory estimateArticle fact-checked
LingBot-VLA 2.0
PermissiveRobbyant (Ant Group) · Robotics · Jul 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 6B
The Qwen3-VL backbone and other dependencies keep their own license terms.
Weights on Hugging FaceArticle fact-checked
Leanstral 1.5
PermissiveMistral AI · Language · Jun 30, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 119B total, 6.5B active (about 66 GB at 4-bit)
NVIDIA STRATA
PermissiveNVIDIA, with U.S. national laboratories · Science · Jun 30, 2026
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- 331.7M
Kimi K2.7-Code
Custom licenseMoonshot AI · Language · Jun 12, 2026
- License
- Modified MIT
- Commercial
- Allowed with conditions
- Parameters
- 1T total, 32B active (about 559 GB at 4-bit)
MIT plus one clause: products over 100M monthly active users or $20M monthly revenue must display Kimi K2.7 Code in the user interface.
Weights on Hugging FaceMemory estimateArticle fact-checked
DiffusionGemma
PermissiveGoogle DeepMind · Language · Jun 10, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 25.2B total, 3.8B active (about 14 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
North Mini Code
PermissiveCohere · Language · Jun 9, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 30B total, about 3B active (about 17 GB at 4-bit)
MiniMax M3
Custom licenseMiniMax · Language · Jun 1, 2026
- License
- MiniMax Community License
- Commercial
- Allowed with conditions
- Parameters
- About 428B total, about 23B active (about 239 GB at 4-bit)
Commercial use needs a Built with MiniMax M3 notice and a notice to MiniMax; written authorization above $20M yearly revenue.
Weights on Hugging FaceMemory estimateArticle fact-checked
NVIDIA Alpamayo 2 Super
PermissiveNVIDIA · Robotics · Jun 1, 2026
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- 34 billion (32B vision-language backbone plus a diffusion Action Expert of about 2B); announced in June 2026 as 32 billion
Inference code is Apache 2.0.
Weights on Hugging FaceArticle fact-checked
NVIDIA Cosmos 3
PermissiveNVIDIA · Robotics · Jun 1, 2026
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- 4B (Edge), 16B (Nano), 64B (Super)
Weights on Hugging FaceArticle fact-checked
GLM-5.2
PermissiveZhipu AI (Z.ai) · Language · Jun 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 744B total, 40B active per Z.ai's model index (Hugging Face counts about 753B, including the multi-token prediction layer) (about 416 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Hy Memory
PermissiveTencent Hunyuan (Tencent Hy) · Language · May 28, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 3B active (Hy-Memory-30B-A3B) (about 1.7 GB at 4-bit)
The page covers a memory system; its Hy-Memory-30B-A3B model, the SDK and the MemHub app are all Apache 2.0.
Weights on Hugging FaceMemory estimateArticle fact-checked
ZAYA1-8B
PermissiveZyphra · Language · May 6, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 8.4B total; 760M active per token (about 4.7 GB at 4-bit)
Mistral Medium 3.5
Custom licenseMistral AI · Language · Apr 28, 2026
- License
- Modified MIT
- Commercial
- Allowed with conditions
- Parameters
- ~128 billion (all active per token) (about 72 GB at 4-bit)
Not licensed to companies with more than $20M monthly revenue, which need a commercial license from Mistral.
DeepSeek V4
PermissiveDeepSeek · Language · Apr 24, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 1.6T total, 49B active (Pro); 284B total, 13B active (Flash) (about 894 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
DeepSeek V4-Flash
PermissiveDeepSeek · Language · Apr 24, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 284B total, 13B active (about 159 GB at 4-bit)
Preview, official 0731 and Vision-Exp checkpoints are all MIT; the hosted V4-Flash API names were retired on September 10, 2026.
Weights on Hugging FaceMemory estimateArticle fact-checked
DeepSeek V4-Pro
PermissiveDeepSeek · Language · Apr 24, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 1.6T total, 49B active (about 894 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Tencent Hunyuan Hy3
PermissiveTencent · Language · Apr 23, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~295B total; ~21B active per token (about 165 GB at 4-bit)
Hy3-preview (April 2026) used the Tencent Hy Community License; the official Hy3 weights (July 2026) are Apache 2.0.
Weights on Hugging FaceMemory estimateArticle fact-checked
Kimi K2.6
Custom licenseMoonshot AI · Language · Apr 20, 2026
- License
- Modified MIT
- Commercial
- Allowed with conditions
- Parameters
- ~1T total; ~32B active per token (about 559 GB at 4-bit)
MIT plus one clause: products over 100M monthly active users or $20M monthly revenue must display Kimi K2.6 in the user interface.
Weights on Hugging FaceMemory estimateArticle fact-checked
NVIDIA Isaac GR00T N1.7
Custom licenseNVIDIA · Robotics · Apr 17, 2026
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
- Parameters
- 3B
Model card and bundled LICENSE give the NVIDIA Open Model License; two subcards still link an older noncommercial license. Code is Apache 2.0.
Weights on Hugging FaceArticle fact-checked
HY-World 2.0
Custom licenseTencent Hunyuan · Vision · Apr 16, 2026
- License
- Tencent Hunyuan Community License
- Commercial
- Allowed with conditions
- Parameters
- About 425M to 80B per component (WorldMirror 2.0 1.2B, WorldStereo 2.0 17B, HY-Pano 2.0 80B)
The license does not apply in the EU, the UK and South Korea, and carries an acceptable use policy.
Lyra 2.0
Non-commercialNVIDIA Spatial Intelligence Lab · Vision · Apr 15, 2026
- License
- NVIDIA Internal Scientific R&D Model License
- Commercial
- Not allowed
Internal research and development only: no distribution, deployment or production use. The code is Apache 2.0.
NVIDIA Ising
PermissiveNVIDIA · Other · Apr 14, 2026
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- 31B (Calibration 1.5); 35B total, 3B active (Calibration 1)
Calibration 1.5 and the ColorCode decoder are OpenMDW 1.1; Calibration 1 and the SurfaceCode decoders use the NVIDIA Open Model License. Decoders are gated.
Article fact-checked
Muse Spark
Partly openMeta Superintelligence Labs · Language · Apr 8, 2026
- License
- Muse Model License
- Commercial
- Allowed with conditions
- Parameters
- 816B total, 42B active (Muse Spark Open) (about 456 GB at 4-bit)
Muse Spark 1.0 to 1.3 are hosted only. Muse Spark Open (October 2026) has gated weights under the Muse Model License.
Memory estimateArticle fact-checked
GLM-5.1
PermissiveZhipu AI (Z.ai) · Language · Apr 7, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 744B total, 40B active per Z.ai's model index (Hugging Face counts about 754B, including the multi-token prediction layer) (about 416 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Gemma 4
PermissiveGoogle DeepMind · Language · Apr 2, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- E2B 5.1B; E4B 8B; 26B A4B 25.2B (3.8B active); 31B 30.7B (about 1.1 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Hy-Embodied
PermissiveTencent (Hunyuan team and Tencent Robotics X) · Robotics · Apr 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 4B (VLA-0.5), 6.2B (RxBrain-1.0), 30B total with about 3B activated (VLM-1.0)
VLM-1.0, RxBrain-1.0 and VLA-0.5 are Apache 2.0. The earlier HY-Embodied-0.5 and 0.5-X use the Tencent HY Community License (excludes EU, UK, South Korea).
Qwen3.6
Partly openAlibaba (Qwen team) · Language · Apr 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- Open: 35B/3B active (MoE), 27B (dense); hosted undisclosed (about 20 GB at 4-bit)
Qwen3.6-35B-A3B and Qwen3.6-27B are Apache 2.0; Qwen3.6-Plus, Flash and Max-Preview are hosted only.
Weights on Hugging FaceMemory estimateArticle fact-checked
Xiaomi MiMo-V2.5
PermissiveXiaomi · Other · Apr 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 310B total, 15B active (standard); 1.02T total, 42B active (Pro)
Article fact-checked
Cohere Transcribe
PermissiveCohere and Cohere Labs · Speech and audio · Mar 26, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 2B (2,065,804,048 in the published checkpoint)
Weights on Hugging FaceArticle fact-checked
MiniMax M2.7
Non-commercialMiniMax · Language · Mar 18, 2026
- License
- MiniMax M2.7 Non-Commercial License
- Commercial
- Not allowed
- Parameters
- ~230B total; ~10B active per token (about 129 GB at 4-bit)
MIT-style text titled Non-Commercial License. Commercial use needs prior written authorization from MiniMax.
Mistral Small 4
PermissiveMistral AI · Language · Mar 16, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 119 billion total; ~6.5 billion active per token (about 66 GB at 4-bit)
Qwen3.5
PermissiveQwen team, Alibaba · Language · Feb 16, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 397B total; about 17B active per token (flagship Qwen3.5-397B-A17B) (about 222 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
GLM-5
PermissiveZhipu AI (Z.ai) · Language · Feb 11, 2026
- License
- MIT
- Commercial
- Allowed
- Parameters
- 744B total; ~40B active per token (about 416 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Qwen3-Coder-Next
PermissiveAlibaba (Qwen team) · Language · Feb 3, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 80B total, about 3B active per token (about 45 GB at 4-bit)
Kimi K2.5
Custom licenseMoonshot AI · Language · Jan 27, 2026
- License
- Modified MIT
- Commercial
- Allowed with conditions
- Parameters
- 1T total; 32B active per token (about 559 GB at 4-bit)
MIT plus one clause: products over 100M monthly active users or $20M monthly revenue must display Kimi K2.5 in the user interface.
Weights on Hugging FaceMemory estimateArticle fact-checked
Jamba2
PermissiveAI21 Labs · Language · Jan 8, 2026
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 3B dense; Mini 52B total, 12B active (about 1.7 GB at 4-bit)
Nemotron 3
Custom licenseNVIDIA · Language · Dec 15, 2025
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
- Parameters
- Nano 30B (3.5B active); Super 120.6B (12.7B); Ultra 550B (55B) (about 17 GB at 4-bit)
Nano and Super use the NVIDIA Nemotron Open Model License. Ultra (June 2026) is under the permissive OpenMDW 1.1.
Memory estimateArticle fact-checked
NVIDIA Isaac GR00T N1.6
Non-commercialNVIDIA (GEAR team) · Robotics · Dec 15, 2025
- License
- NVIDIA License (non-commercial)
- Commercial
- Not allowed
- Parameters
- 3B
Weights are limited to noncommercial research or evaluation (except NVIDIA and affiliates); the code is Apache 2.0.
Mistral Large 3
PermissiveMistral AI · Language · Dec 2, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~675B total; ~41B active per token (vision encoder ~2.5B) (about 377 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
FLUX.2
Partly openBlack Forest Labs · Image · Nov 25, 2025
- License
- FLUX Non-Commercial License
- Commercial
- Not allowed
- Parameters
- ~32B (dev, pro, flex, max); klein 4B and 9B
[dev] and klein 9B are non-commercial; klein 4B is Apache 2.0; [pro], [flex] and [max] are API only. BFL sells commercial licenses.
Weights on Hugging FaceArticle fact-checked
OLMo 3
PermissiveAllen Institute for AI (Ai2) · Language · Nov 20, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 7B and 32B (about 3.9 GB at 4-bit)
Omnilingual ASR
PermissiveMeta FAIR · Speech and audio · Nov 10, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 0.32B to 7.81B
The commissioned Omnilingual ASR Corpus is published under CC BY 4.0; the rest of the training data is not.
Article fact-checked
Kimi K2 Thinking
Custom licenseMoonshot AI · Language · Nov 6, 2025
- License
- Modified MIT
- Commercial
- Allowed with conditions
- Parameters
- About 1 trillion total; about 32 billion active per token (about 559 GB at 4-bit)
MIT plus an attribution clause for products over 100M monthly active users or $20M monthly revenue.
Weights on Hugging FaceMemory estimateArticle fact-checked
Kimi Linear
PermissiveMoonshot AI · Language · Oct 30, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 48B total, 3B active (about 27 GB at 4-bit)
MiniMax M2
Custom licenseMiniMax · Language · Oct 27, 2025
- License
- Modified MIT
- Commercial
- Allowed with conditions
- Parameters
- 230B total; ~10B active per token (about 129 GB at 4-bit)
MIT at launch, amended Nov 13, 2025: products over 100M monthly users or $30M annual revenue must display MiniMax M2 in the interface.
Weights on Hugging FaceMemory estimateArticle fact-checked
DeepSeek-OCR
PermissiveDeepSeek · Vision · Oct 20, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- About 3B decoder (about 570M activated per token) plus about 380M encoder
Ling-1T
PermissiveAnt Group (Inclusion AI) · Language · Oct 9, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- ~1 trillion total; 51B active per token (about 559 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Jamba Reasoning 3B
PermissiveAI21 Labs · Language · Oct 8, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~3B (about 1.7 GB at 4-bit)
IBM Granite 4.0
PermissiveIBM · Language · Oct 2, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~3B (Micro) to ~32B total with ~9B active (H-Small) (about 1.7 GB at 4-bit)
Ring-1T
PermissiveinclusionAI (Ant Group) · Language · Oct 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- ~1T total; ~50B active per token (about 559 GB at 4-bit)
GLM-4.6
PermissiveZhipu AI (Z.ai) · Language · Sep 30, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- ~357B total; ~32B active per token (about 199 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
DeepSeek V3.2
PermissiveDeepSeek · Language · Sep 29, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 671B total; 37B active per token (about 375 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Hunyuan Image 3.0
Custom licenseTencent · Image · Sep 28, 2025
- License
- Tencent Hunyuan Community License
- Commercial
- Allowed with conditions
- Parameters
- ~80B total; ~13B active per token; 64 experts
Excludes the EU, UK and South Korea; above 100 million monthly active users a separate license is required.
Qwen3-VL
PermissiveQwen team, Alibaba Cloud · Language · Sep 23, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- Dense 2B, 4B, 8B, 32B; MoE 30B-A3B and 235B-A22B (about 1.1 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
EmbeddingGemma
Custom licenseGoogle DeepMind · Language · Sep 4, 2025
- License
- Gemma Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- ~308M (~100M backbone, ~200M embedding table) (about 0.2 GB at 4-bit)
Gated on Hugging Face: accept the Gemma Terms of Use first. Commercial use allowed subject to the prohibited use policy.
Apertus
PermissiveSwiss AI Initiative (EPFL, ETH Zurich, CSCS) · Language · Sep 2, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 8B and 70B (70B reported as 71B); base and Instruct variants (about 4.5 GB at 4-bit)
LongCat-Flash
PermissiveMeituan (LongCat team) · Language · Sep 1, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- ~560B total; 18.6B-31.3B active per token (avg ~27B) (about 313 GB at 4-bit)
Qwen3-Next
PermissiveQwen team, Alibaba Cloud · Language · Sep 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 80 billion total; ~3 billion active per token (about 45 GB at 4-bit)
Qwen3-Omni
PermissiveQwen team, Alibaba Cloud · Language · Sep 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 30B total (Thinker); ~3B active per token (about 17 GB at 4-bit)
Command A Reasoning
Non-commercialCohere and Cohere Labs · Language · Aug 21, 2025
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 111 billion (about 62 GB at 4-bit)
Gated on Hugging Face (contact details required). Commercial use only through Cohere.
Jet-Nemotron
Non-commercialNVIDIA Research · Language · Aug 21, 2025
- License
- NVIDIA License (non-commercial)
- Commercial
- Not allowed
- Parameters
- ~2B and ~4B variants (about 1.1 GB at 4-bit)
Weights may be used for research or evaluation only; the code is Apache 2.0.
rBio (CZI)
Non-commercialChan Zuckerberg Initiative (CZI) · Science · Aug 21, 2025
- License
- Qwen Research License
- Commercial
- Not allowed
- Parameters
- 3 billion (base model)
Built on Qwen2.5-3B-Instruct, so the Qwen Research License applies; CZI's own changes are MIT. Weights are on AWS S3, not Hugging Face.
DeepSeek V3.1
PermissiveDeepSeek · Language · Aug 19, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 671B total, 37B active per token (685B checkpoint incl. ~14B MTP module) (about 375 GB at 4-bit)
Nemotron Nano 2
Custom licenseNVIDIA · Language · Aug 18, 2025
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
- Parameters
- ~9B (flagship, pruned and distilled from a 12B base) (about 5.0 GB at 4-bit)
Most of the pre-training corpus is published as Nemotron-Pre-Training-Dataset-v1 under CC BY 4.0.
DINOv3
Custom licenseMeta AI (FAIR) · Vision · Aug 14, 2025
- License
- DINOv3 License
- Commercial
- Allowed with conditions
- Parameters
- 21M to 6.7B (ViT-7B/16 largest)
Gated on Hugging Face (manual approval). Commercial use allowed; the license bars ITAR-controlled and trade-control-prohibited end uses.
gpt-oss
PermissiveOpenAI · Language · Aug 5, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 120b: 117B total, 5.1B active; 20b: 21B total, 3.6B active (about 67 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Qwen-Image
PermissiveQwen team, Alibaba (Tongyi Lab) · Other · Aug 4, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~20 billion (MMDiT backbone)
Weights on Hugging FaceArticle fact-checked
Hermes 4
Custom licenseNous Research · Language · Aug 2025
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 14B, 70B, 405B (about 7.8 GB at 4-bit)
70B and 405B (Llama 3.1 base) use the Llama 3.1 Community License; the 14B (Qwen3 base) is Apache 2.0.
FoxBrain
Partly openHon Hai Research Institute (Foxconn) · Language · Jul 31, 2025
- License
- FoxBrain Model Usage and License Agreement
- Commercial
- Not allowed
- Parameters
- 70B (v1.x) (about 39 GB at 4-bit)
v1.2 (70B) is manually gated for academic and research organizations; the v1.0 preview is partner-only and no FoxBrain 2.0 weights are published.
Weights on Hugging FaceMemory estimateArticle fact-checked
Step-3
PermissiveStepFun · Language · Jul 31, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~321B total; ~38B active per token (about 179 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
GLM-4.5
PermissiveZhipu AI (Z.ai) · Language · Jul 28, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 355B total, 32B active; Air: 106B total, 12B active (about 198 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
HunyuanWorld 1.0
Custom licenseTencent (Hunyuan team) · Vision · Jul 26, 2025
- License
- Tencent Hunyuan Community License
- Commercial
- Allowed with conditions
License excludes the EU, UK and South Korea; licensees above 1 million monthly active users must request a separate license from Tencent.
Qwen3-Coder
PermissiveAlibaba Qwen team (Tongyi Lab) · Language · Jul 22, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 480B total; 35B active per token (flagship) (about 268 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Voxtral
PermissiveMistral AI · Speech and audio · Jul 15, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- Voxtral Small 24.3B; Voxtral Mini 4.7B
Kimi K2
Custom licenseMoonshot AI · Language · Jul 11, 2025
- License
- Modified MIT
- Commercial
- Allowed with conditions
- Parameters
- 1.04T total; 32.6B active per token (about 581 GB at 4-bit)
MIT plus one clause: products over 100M monthly active users or $20M monthly revenue must display Kimi K2 in the user interface.
Weights on Hugging FaceMemory estimateArticle fact-checked
SmolLM 3
PermissiveHugging Face · Language · Jul 8, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 3B (about 1.7 GB at 4-bit)
Phi-4-mini-flash-reasoning
PermissiveMicrosoft · Language · Jul 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 3.8 billion (about 2.1 GB at 4-bit)
UMA (Universal Model for Atoms)
Custom licenseFAIR Chemistry team, Meta AI · Science · Jun 30, 2025
- License
- FAIR Chemistry License v1
- Commercial
- Allowed with conditions
- Parameters
- UMA-S ~150M, UMA-M ~1.4B, UMA-L ~700M (total)
Gated on Hugging Face with manual approval; commercial use allowed under an acceptable use policy that excludes some jurisdictions.
AlphaGenome
Non-commercialGoogle DeepMind · Science · Jun 25, 2025
- License
- AlphaGenome Model Terms
- Commercial
- Not allowed
Weights on Kaggle and Hugging Face after accepting non-commercial model terms; the code is Apache 2.0. Commercial use goes through Google Cloud.
State (Arc Institute virtual cell model)
Non-commercialArc Institute · Science · Jun 23, 2025
- License
- Arc Research Institute State Model Non-Commercial License
- Commercial
- Not allowed
- Parameters
- SE-600M checkpoint about 0.7B
Weights are for non-commercial entities and purposes only; the code is CC BY-NC-SA 4.0.
MiniMax M1
PermissiveMiniMax · Language · Jun 16, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 456B total; 45.9B active per token (about 255 GB at 4-bit)
V-JEPA 2
PermissiveMeta AI · Robotics · Jun 11, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 1.2B (world-model configuration; encoder over 1B)
Article fact-checked
Magistral
Partly openMistral AI · Language · Jun 10, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 24B (Magistral Small); unpublished (Magistral Medium) (about 13 GB at 4-bit)
Magistral Small (24B) is open under Apache 2.0. Magistral Medium is available only through Mistral's API and products.
SmolVLA
PermissiveHugging Face (LeRobot team) · Robotics · Jun 3, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~450M total (~100M action expert)
Weights on Hugging FaceArticle fact-checked
Boltz-2
PermissiveMIT (Jameel Clinic, CSAIL) and Recursion · Science · Jun 2025
- License
- MIT
- Commercial
- Allowed
Hunyuan-A13B
Custom licenseTencent (Hunyuan team) · Language · Jun 2025
- License
- Tencent Hunyuan Community License
- Commercial
- Allowed with conditions
- Parameters
- ~80B total; ~13B active per token (about 45 GB at 4-bit)
Excludes the EU, UK and South Korea; above 100 million monthly active users a separate license is required.
Qwen3 Embedding
PermissiveAlibaba Qwen team · Language · Jun 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 0.6B, 4B, 8B (about 0.3 GB at 4-bit)
Devstral
Partly openMistral AI with All Hands AI · Language · May 21, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 24 billion (Small); Medium not disclosed (about 13 GB at 4-bit)
Devstral Small (2505, 2507, Small 2) is Apache 2.0; Devstral Medium is API only; Devstral 2 123B uses a modified MIT with a revenue cap.
Falcon-H1
Custom licenseTechnology Innovation Institute (TII), Abu Dhabi · Language · May 20, 2025
- License
- TII Falcon License
- Commercial
- Allowed with conditions
- Parameters
- Six sizes: 0.5B, 1.5B, 1.5B-Deep, 3B, 7B, 34B (about 0.3 GB at 4-bit)
Apache-2.0-based TII Falcon LLM License with an acceptable use policy.
Gemma 3n
Custom licenseGoogle DeepMind · Language · May 20, 2025
- License
- Gemma Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- 5B (E2B) and 8B (E4B) raw; about 2B and 4B effective (about 2.8 GB at 4-bit)
Hugging Face repositories are gated behind acceptance of the Gemma terms.
MedGemma
Custom licenseGoogle · Science · May 20, 2025
- License
- Health AI Developer Foundations Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- 4B and 27B
Gated: the HAI-DEF terms must be accepted. Use is bound by a prohibited use policy and health regulatory requirements.
Wan 2.1-VACE
PermissiveAlibaba (Tongyi Lab) · Video · May 14, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 1.3B and 14B (two variants)
Isaac GR00T N1.5
Non-commercialNVIDIA · Robotics · May 2025
- License
- NVIDIA License (non-commercial)
- Commercial
- Not allowed
- Parameters
- ~3B (GR00T-N1.5-3B)
Research or evaluation use only. N1.7 weights use the NVIDIA Open Model License; repository code is Apache 2.0.
NVIDIA Cosmos Reason
Custom licenseNVIDIA · Robotics · May 2025
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
- Parameters
- 7B (Cosmos-Reason1); 2B to 32B (Cosmos-Reason2)
Cosmos-Reason1-7B and the Cosmos-Reason2 models are downloadable; the 56B Reason1 model was described in the paper only.
Phi-4 Reasoning
PermissiveMicrosoft · Language · Apr 30, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 14 billion (about 7.8 GB at 4-bit)
Qwen3
PermissiveQwen Team, Alibaba Cloud · Language · Apr 29, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 0.6B to 235B (flagship 235B total; 22B active) (about 0.3 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
π0.5
Custom licensePhysical Intelligence · Robotics · Apr 22, 2025
- License
- Gemma Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- ~3.3B (~3B PaliGemma backbone + ~300M action expert)
Released in openpi in September 2025 (base, LIBERO and DROID checkpoints). openpi is Apache 2.0 and ships the Gemma Terms for the PaliGemma-derived weights.
Article fact-checked
InternVL3
Custom licenseOpenGVLab (Shanghai AI Laboratory) · Language · Apr 11, 2025
- License
- Qwen License (Tongyi Qianwen)
- Commercial
- Allowed with conditions
- Parameters
- Seven dense sizes, 938M to 78.41B (about 0.5 GB at 4-bit)
Applies to the flagship InternVL3-78B (Qwen2.5-72B base); the 1B to 38B sizes are tagged Apache 2.0 and the 9B is MIT.
Llama 4
Custom licenseMeta · Language · Apr 5, 2025
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- Scout: 109B total, 17B active; Maverick: ~400B total, 17B active (about 61 GB at 4-bit)
Scout and Maverick are downloadable (gated); Behemoth was never released. The license grant excludes EU-based individuals and companies.
Weights on Hugging FaceMemory estimateArticle fact-checked
Llama 4 Scout and Maverick
Custom licenseMeta · Language · Apr 5, 2025
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- Scout: ~109B total, 17B active (16 experts); Maverick: ~400B total, 17B active (128 experts) (about 61 GB at 4-bit)
Gated download. License grant excludes individuals and companies based in the EU; 700M monthly active user clause.
Apriel (ServiceNow)
PermissiveServiceNow (AI Research / SLAM lab) · Language · Apr 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- ~4.8B to ~15B (about 2.7 GB at 4-bit)
HiDream
PermissiveHiDream.ai · Image · Apr 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- About 17 billion
Nemotron-H
Non-commercialNVIDIA · Language · Apr 2025
- License
- NVIDIA Internal Scientific R&D Model License
- Commercial
- Not allowed
- Parameters
- 8B, 47B, and 56B (base models) (about 4.5 GB at 4-bit)
Research and development use only under NVIDIA's research license.
TxGemma
Custom licenseGoogle · Science · Mar 25, 2025
- License
- Health AI Developer Foundations Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- 2B, 9B, 27B
Gated: the HAI-DEF terms must be accepted. Use is bound by a prohibited use policy and health regulatory requirements.
Llama Nemotron
Custom licenseNvidia · Language · Mar 18, 2025
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
- Parameters
- 8B (Nano); 49B (Super); 253B (Ultra) (about 4.5 GB at 4-bit)
Also subject to the Llama 3.1 or 3.3 Community License. The post-training dataset is published; the Llama base data is not.
NVIDIA Isaac GR00T N1
Non-commercialNVIDIA (GEAR research lab) · Robotics · Mar 18, 2025
- License
- NVIDIA License (non-commercial)
- Commercial
- Not allowed
- Parameters
- ~2B (base N1)
N1, N1.5 and N1.6 weights are for research or evaluation only; N1.7 uses the NVIDIA Open Model License. Code is Apache 2.0.
RWKV-7 (Goose)
PermissiveRWKV Project (Bo Peng et al., with EleutherAI and Recursal AI) · Language · Mar 18, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 0.19B to 2.9B (World v3 series); later 7.2B G1 checkpoint (about 0.1 GB at 4-bit)
ERNIE 4.5
PermissiveBaidu · Language · Mar 16, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 0.3B dense up to 424B total (47B active) across ten variants (about 0.2 GB at 4-bit)
Cohere Command A
Non-commercialCohere · Language · Mar 13, 2025
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 111B (about 62 GB at 4-bit)
Gated on Hugging Face (contact details required). Commercial use only through Cohere's API, cloud partners or a Cohere license.
Sesame CSM
Partly openSesame AI · Speech and audio · Mar 13, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 1B backbone + 100M audio decoder (CSM-1B)
Only CSM-1B is released (Apache 2.0, click-through gate on Hugging Face); the larger 3B and 8B variants are not public.
Gemma 3
Custom licenseGoogle DeepMind · Language · Mar 12, 2025
- License
- Gemma Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- 1B, 4B, 12B, 27B (270M added August 2025) (about 0.6 GB at 4-bit)
Commercial use allowed under the Gemma Terms of Use and its prohibited use policy; Hugging Face access requires accepting the terms.
Weights on Hugging FaceMemory estimateArticle fact-checked
Skywork-R1V
PermissiveSkywork AI (Kunlun Tech) · Language · Mar 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 38B (each open generation) (about 21 GB at 4-bit)
Phi-4-mini
PermissiveMicrosoft · Language · Feb 26, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 3.8 billion (about 2.1 GB at 4-bit)
Wan 2.1
PermissiveAlibaba (Tongyi Wanxiang team) · Video · Feb 25, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 1.3B and 14B variants
Weights on Hugging FaceArticle fact-checked
Evo 2 (genomic model)
PermissiveArc Institute, NVIDIA, Stanford, and collaborators · Science · Feb 19, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 40 billion (largest); also 7B and 1B
Step-Video
PermissiveStepFun (Shanghai Jieyue Xingchen) · Video · Feb 17, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 30B (DiT backbone)
LLaDA (Large Language Diffusion)
PermissiveRenmin University of China and Ant Group · Language · Feb 14, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 8B (LLaDA 8B); up to 100B (LLaDA 2.0 flash) (about 4.5 GB at 4-bit)
LLaDA 8B Base/Instruct and LLaDA 1.5 are MIT; LLaDA 2.0 mini and flash (inclusionAI) are Apache 2.0.
Goedel-Prover
PermissivePrinceton Language and Intelligence, with Tsinghua, Amazon, Meta FAIR, Numina · Language · Feb 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 7B (V1); 8B and 32B (V2) (about 3.9 GB at 4-bit)
V2 weights (8B, 32B) are Apache 2.0; V1 weights are MIT.
NVIDIA COMPASS
Custom licenseNVIDIA · Robotics · Feb 2025
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
The checkpoint download requires accepting NVIDIA's terms on Hugging Face; the code is Apache 2.0.
Weights on Hugging FaceArticle fact-checked
olmOCR
PermissiveAllen Institute for AI (Ai2) · Language · Feb 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 7B (about 3.9 GB at 4-bit)
Qwen2.5-VL
PermissiveAlibaba Cloud (Qwen team) · Vision · Jan 26, 2025
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 3B, 7B, 32B, 72B
7B and 32B are Apache 2.0. Exceptions: 3B is under the Qwen Research License (non-commercial); 72B is under the Qwen License.
Article fact-checked
DeepSeek-R1
PermissiveDeepSeek · Language · Jan 20, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 671B total; 37B active per token (about 375 GB at 4-bit)
R1, R1-Zero and R1-0528 weights are MIT. Training data and the training framework were not released.
Weights on Hugging FaceMemory estimateArticle fact-checked
DeepSeek-R1-Distill
PermissiveDeepSeek · Language · Jan 20, 2025
- License
- MIT
- Commercial
- Allowed
- Parameters
- 1.5B, 7B, 8B, 14B, 32B, and 70B variants (about 0.8 GB at 4-bit)
All six are tagged MIT; the Qwen bases are Apache 2.0, and the two Llama-based variants also carry Llama 3.1 or 3.3 license terms.
MatterGen
PermissiveMicrosoft Research, AI for Science · Science · Jan 16, 2025
- License
- MIT
- Commercial
- Allowed
MiniMax-Text-01
Custom licenseMiniMax · Language · Jan 14, 2025
- License
- MiniMax Model License
- Commercial
- Allowed with conditions
- Parameters
- 456B total; 45.9B active per token (about 255 GB at 4-bit)
License revised Feb 20, 2025. Commercial use is allowed subject to a prohibited-uses policy that bars military use.
NVIDIA Cosmos
PermissiveNVIDIA · Robotics · Jan 6, 2025
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- 4B to 14B (Cosmos 1.0 models)
Cosmos 3 (2026) is under OpenMDW 1.1. Cosmos 1.0 to 2.5 stay under the NVIDIA Open Model License, with gated downloads.
Article fact-checked
DeepSeek V3
Custom licenseDeepSeek · Language · Dec 26, 2024
- License
- DeepSeek Model License
- Commercial
- Allowed with conditions
- Parameters
- 671B total (about 685B with MTP module); 37B active per token (about 375 GB at 4-bit)
Original V3 and V3-Base weights use the DeepSeek Model License; the V3-0324 weights (March 2025) are MIT.
Weights on Hugging FaceMemory estimateArticle fact-checked
QvQ
Partly openQwen team, Alibaba · Language · Dec 25, 2024
- License
- Qwen License (Tongyi Qianwen)
- Commercial
- Allowed with conditions
- Parameters
- 72B (QVQ-72B-Preview) (about 40 GB at 4-bit)
QVQ-72B-Preview weights are under the Qwen license; QVQ-Max is offered only through Qwen Chat and Alibaba Cloud.
Falcon 3
Custom licenseTechnology Innovation Institute (TII) · Language · Dec 17, 2024
- License
- TII Falcon License
- Commercial
- Allowed with conditions
- Parameters
- 1B, 3B, 7B, 10B (transformer); 7B (Mamba) (about 0.6 GB at 4-bit)
Apache-2.0-derived license with an acceptable use policy; offering shared hosted instances needs a separate TII agreement.
Byte Latent Transformer
Non-commercialMeta AI (FAIR) · Language · Dec 13, 2024
- License
- FAIR Noncommercial Research License
- Commercial
- Not allowed
- Parameters
- Up to 8B trained; 1B and 7B checkpoints released (about 4.5 GB at 4-bit)
BLT 1B and 7B checkpoints are on Hugging Face behind an access request; the code is CC BY-NC 4.0.
DeepSeek-VL2
Custom licenseDeepSeek (DeepSeek-AI) · Vision · Dec 13, 2024
- License
- DeepSeek Model License
- Commercial
- Allowed with conditions
- Parameters
- 3B/16B/27B total; 1.0B/2.8B/4.5B active
Code is MIT; weights use the DeepSeek Model License, which allows commercial use subject to use restrictions.
Meta Motivo
Non-commercialMeta AI (FAIR) · Robotics · Dec 12, 2024
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- ~24.5M (S variants); ~288M (metamotivo-M-1)
Phi-4
PermissiveMicrosoft Research · Language · Dec 12, 2024
- License
- MIT
- Commercial
- Allowed
- Parameters
- 14 billion (about 14.7B with embeddings) (about 7.8 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Llama 3.3
Custom licenseMeta · Language · Dec 6, 2024
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 70 billion (about 39 GB at 4-bit)
Gated download. Acceptable use policy applies; over 700M monthly active users needs a separate Meta license.
Weights on Hugging FaceMemory estimateArticle fact-checked
GenCast
PermissiveGoogle DeepMind · Science · Dec 4, 2024
- License
- CC BY 4.0
- Commercial
- Allowed
Weights were relicensed from CC BY-NC-SA 4.0 to CC BY 4.0 on August 6, 2026 to permit commercial use; code is Apache 2.0.
HunyuanVideo
Custom licenseTencent · Video · Dec 3, 2024
- License
- Tencent Hunyuan Community License
- Commercial
- Allowed with conditions
- Parameters
- 13B (1.0); 8.3B (1.5)
License does not apply in the EU, UK or South Korea; over 100M monthly active users needs a separate Tencent license.
Weights on Hugging FaceArticle fact-checked
QwQ
PermissiveAlibaba Cloud Qwen team · Language · Nov 28, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 32.5B (dense) (about 18 GB at 4-bit)
OLMo 2
PermissiveAllen Institute for AI (Ai2) · Language · Nov 26, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 7B, 13B, 32B (about 3.9 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
LTX-Video
Custom licenseLightricks · Video · Nov 21, 2024
- License
- LTXV Open Weights License
- Commercial
- Allowed with conditions
- Parameters
- ~1.9B (base); 13B (LTXV-13B)
Entities with annual revenue of $10M or more need a paid commercial license. Versions before 0.9.6 shipped under earlier licenses (0.9: RAIL-M).
Tülu 3
Custom licenseAllen Institute for AI (Ai2) · Other · Nov 21, 2024
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 8B, 70B, 405B
Inherits the Llama 3.1 Community License (acceptable use policy, 700M monthly user threshold); the card also cites Gemma and Qwen terms.
Article fact-checked
Pixtral Large
Non-commercialMistral AI · Language · Nov 18, 2024
- License
- Mistral Research License
- Commercial
- Not allowed
- Parameters
- ~124 billion total (about 69 GB at 4-bit)
Research and non-commercial use only; commercial use needs a separate Mistral Commercial License.
Boltz
PermissiveMIT (CSAIL and Jameel Clinic) · Science · Nov 1, 2024
- License
- MIT
- Commercial
- Allowed
SmolLM 2
PermissiveHugging Face · Language · Nov 1, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 135M, 360M, 1.7B (three sizes) (about 0.1 GB at 4-bit)
Hunyuan 3D
Partly openTencent · Other · Nov 2024
- License
- Tencent Hunyuan Community License
- Commercial
- Allowed with conditions
- Parameters
- Up to ~10B shape model (2.5); 3.3B shape + 2B paint in 2.1
Weights open up to 2.1; 2.5, 3.0 and 3.1 are hosted only. Excludes EU, UK and South Korea; above 1M monthly users needs a separate license.
Weights on Hugging FaceArticle fact-checked
Marco-o1
PermissiveAlibaba International Digital Commerce (MarcoPolo team) · Language · Nov 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 7.62B (7,615,616,512 in the published checkpoint) (about 4.3 GB at 4-bit)
π0
Custom licensePhysical Intelligence · Robotics · Oct 31, 2024
- License
- Gemma Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- ~3.3B total (3B VLM backbone + ~300M action expert)
openpi is Apache 2.0 and also ships the Gemma Terms of Use, which cover the PaliGemma-derived weights. The pi training dataset is proprietary.
Article fact-checked
GLM-4-Voice
Custom licenseZhipu AI and Tsinghua University KEG · Speech and audio · Oct 25, 2024
- License
- glm-4-voice License
- Commercial
- Allowed with conditions
- Parameters
- 9B (GLM-4-9B backbone)
Free for academic research; commercial use requires registration with Zhipu AI and a Built with glm-4 notice. Code is Apache 2.0.
Mochi 1
PermissiveGenmo Inc. · Video · Oct 22, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 10 billion (DiT) plus 362 million (VAE)
Stable Diffusion 3.5
Custom licenseStability AI · Image · Oct 22, 2024
- License
- Stability AI Community License
- Commercial
- Allowed with conditions
- Parameters
- 8.1B (Large); ~2.5B (Medium)
Free for commercial use under $1M annual revenue, enterprise license above that; click-through gate on Hugging Face.
SpiRit-LM
Non-commercialMeta AI (FAIR) · Speech and audio · Oct 18, 2024
- License
- FAIR Noncommercial Research License
- Commercial
- Not allowed
- Parameters
- 7B (Base and Expressive versions)
Base and Expressive 7B weights are released for noncommercial research only, with inference code.
DeepSeek Janus
Custom licenseDeepSeek · Vision · Oct 17, 2024
- License
- DeepSeek Model License
- Commercial
- Allowed with conditions
- Parameters
- 1.3B (Janus, JanusFlow); 1B and 7B (Janus-Pro)
Hub tag reads MIT, but the model cards put the weights under the DeepSeek Model License; only the code is MIT.
Ministral 3B / 8B
Partly openMistral AI · Language · Oct 16, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~3B and ~8B (8,019,808,256 for the 8B) (about 1.7 GB at 4-bit)
Ministral 3 (3B, 8B, 14B) is Apache 2.0. Ministral 8B 2410 is under the Mistral Research License and Ministral 3B 2410 was API only.
F5-TTS
Non-commercialShanghai Jiao Tong University X-LANCE Lab, University of Cambridge, Geely · Speech and audio · Oct 9, 2024
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 335.8 million
Code is MIT; the official checkpoints are CC BY-NC 4.0 because of the Emilia training data.
Llama-3.1-Nemotron-70B-Instruct
Custom licenseNVIDIA · Language · Oct 2024
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
- Parameters
- 70B (about 39 GB at 4-bit)
The Llama 3.1 Community License also applies, and the Hugging Face tag is llama3.1.
Llama 3.2
Custom licenseMeta · Language · Sep 25, 2024
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 1.23B, 3.21B (text-only); 11B, 90B (vision) (about 0.7 GB at 4-bit)
Gated download. For the 11B and 90B vision models the license grant excludes individuals and companies based in the EU.
Memory estimateArticle fact-checked
Llama 3.2 Vision
Custom licenseMeta AI · Language · Sep 25, 2024
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 11B (10.6B actual) and 90B (88.8B actual) (about 6.1 GB at 4-bit)
Gated download. The license grant excludes individuals and companies based in the EU (end users of products are exempt).
Weights on Hugging FaceMemory estimateArticle fact-checked
Molmo
PermissiveAllen Institute for AI (Ai2) · Vision · Sep 25, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 72B, 7B (x2), MolmoE-1B (~1.5B active, 7.2B total)
Training code is on GitHub and the PixMo training datasets are published under ODC-BY 1.0.
Qwen2.5
PermissiveAlibaba Cloud Qwen team · Language · Sep 19, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- Seven dense sizes from 0.49B to 72.7B (about 0.3 GB at 4-bit)
Five of seven sizes are Apache 2.0; the 72B is under the Qwen License (conditions on commercial use) and the 3B is research only.
Weights on Hugging FaceMemory estimateArticle fact-checked
Qwen2.5-Math
PermissiveQwen team, Alibaba Cloud · Language · Sep 19, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 1.5B, 7B and 72B (about 0.8 GB at 4-bit)
The 1.5B and 7B models are Apache 2.0; the 72B models and RM-72B are under the Qwen License (conditions on commercial use).
Weights on Hugging FaceMemory estimateArticle fact-checked
Jina Embeddings v3
Non-commercialJina AI · Language · Sep 18, 2024
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 570 million (about 0.3 GB at 4-bit)
NVLM
Non-commercialNVIDIA · Language · Sep 17, 2024
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- ~79B (NVLM-D-72B: 72B LLM plus 6B vision encoder) (about 44 GB at 4-bit)
Pixtral
PermissiveMistral AI · Language · Sep 11, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~12.4B (12B: 12B decoder + 400M encoder); 124B (Large) (about 6.9 GB at 4-bit)
Applies to Pixtral 12B. Pixtral Large (124B) is under the Mistral Research License: non-commercial unless a commercial license is bought.
Memory estimateArticle fact-checked
Chai-1
PermissiveChai Discovery · Science · Sep 9, 2024
- License
- Apache 2.0
- Commercial
- Allowed
Apache 2.0 for code and weights since November 26, 2024 (research-only license before that). Inference code only; no training code.
OLMoE
PermissiveAllen Institute for AI, with Contextual AI · Language · Sep 3, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 6.9B total; ~1.3B active per token (about 3.9 GB at 4-bit)
Hailuo AI
Partly openMiniMax · Video · Sep 2024
- License
- MiniMax H3 Community License Agreement
- Commercial
- Allowed with conditions
Only MiniMax H3 weights are public (Aug 2026), in part; earlier Hailuo models are hosted only. License excludes the EU, UK, South Korea and US.
Article fact-checked
Qwen2.5-Coder
PermissiveQwen team, Alibaba · Language · Sep 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- Six sizes, 0.5B to 32B (flagship 32.5B) (about 0.3 GB at 4-bit)
Five sizes (0.5B, 1.5B, 7B, 14B, 32B) are Apache 2.0; the 3B is under the Qwen Research License, which is non-commercial.
Weights on Hugging FaceMemory estimateArticle fact-checked
Qwen2-VL
PermissiveAlibaba Cloud (Qwen team) · Vision · Aug 29, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 2B, 7B, 72B
2B and 7B are Apache 2.0. Exception: the 72B weights use the Qwen (Tongyi Qianwen) License, which sets conditions on commercial use.
Article fact-checked
Sapiens (computer vision)
Non-commercialMeta (Reality Labs) · Vision · Aug 22, 2024
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 0.3B, 0.6B, 1B, 2B
Code and weights of the 2024 release are CC BY-NC 4.0; the 2026 successor Sapiens2 uses its own Sapiens2 License.
Show-o
PermissiveShow Lab, National University of Singapore (with ByteDance co-authors) · Other · Aug 21, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 1.3B
GitHub LICENSE and the 512x512 and Show-o2 checkpoints are Apache 2.0; the first 256x256 Hub checkpoints are tagged MIT.
Qwen2-Math
PermissiveQwen team, Alibaba · Language · Aug 8, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 1.5B, 7B, and 72B (base and Instruct) (about 0.8 GB at 4-bit)
The 1.5B and 7B sizes are Apache 2.0; Qwen2-Math-72B is under the Tongyi Qianwen License (conditions on commercial use).
CogVideoX
Custom licenseZhipu AI and Tsinghua University (THUDM) · Video · Aug 6, 2024
- License
- CogVideoX License
- Commercial
- Allowed with conditions
- Parameters
- 2B or 5B (by variant)
CogVideoX-2B is Apache 2.0; 5B and 1.5 use the CogVideoX License: commercial use needs free registration, capped at 1M visits a month.
Flux (text-to-image model)
Partly openBlack Forest Labs · Image · Aug 1, 2024
- License
- FLUX Non-Commercial License
- Commercial
- Not allowed
- Parameters
- 12B (FLUX.1); 32B (FLUX.2); 4B and 9B (FLUX.2 klein)
[schnell] and FLUX.2 klein 4B are Apache 2.0; [dev] and klein 9B are non-commercial; FLUX 3 Klein and Action use the FLUX Kommunity License; pro tiers API only.
Article fact-checked
FLUX.1
Partly openBlack Forest Labs · Image · Aug 1, 2024
- License
- FLUX Non-Commercial License
- Commercial
- Not allowed
- Parameters
- ~12B
[schnell] is Apache 2.0; [dev], Kontext [dev], Krea [dev] and the Tools are non-commercial; [pro] is API only. BFL sells commercial licenses.
Weights on Hugging FaceArticle fact-checked
Qwen2-Audio
PermissiveQwen team, Alibaba Cloud · Speech and audio · Aug 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 8.2B
ShieldGemma
Custom licenseGoogle · Other · Jul 31, 2024
- License
- Gemma Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- 2B, 9B, 27B (text); 4B (ShieldGemma 2)
Gated: Google's Gemma terms must be accepted before download. Covers ShieldGemma and ShieldGemma 2.
SAM 2
PermissiveMeta AI (FAIR) · Vision · Jul 29, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- Tiny 38.9M, Small 46M, Base+ 80.8M, Large 224.4M
Llama 3.1
Custom licenseMeta · Language · Jul 23, 2024
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 8B, 70B, and 405B (about 4.5 GB at 4-bit)
Gated download. Acceptable use policy applies; over 700M monthly active users needs a separate Meta license.
Weights on Hugging FaceMemory estimateArticle fact-checked
Mistral NeMo
PermissiveMistral AI, with NVIDIA · Language · Jul 18, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 12B (about 6.7 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
SmolLM
PermissiveHugging Face · Language · Jul 16, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 135M, 360M, and 1.7B variants (about 0.1 GB at 4-bit)
Moshi
PermissiveKyutai · Speech and audio · Jul 3, 2024
- License
- CC BY 4.0
- Commercial
- Allowed
- Parameters
- 7.7B total (Helium backbone is 7B)
CosyVoice
PermissiveTongyi SpeechTeam, Alibaba Group (FunAudioLLM project) · Speech and audio · Jul 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 300M (v1); 0.5B (CosyVoice 2 and the open CosyVoice 3 checkpoint); 1.5B language model in the CosyVoice 3 paper
Gemma 2
Custom licenseGoogle DeepMind · Language · Jun 27, 2024
- License
- Gemma Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- 2.6B, 9B, 27B (about 1.5 GB at 4-bit)
Commercial use allowed under the Gemma Terms of Use and its prohibited use policy; Hugging Face access requires accepting the terms.
Weights on Hugging FaceMemory estimateArticle fact-checked
LLM Compiler (Meta)
Custom licenseMeta AI · Language · Jun 27, 2024
- License
- Meta LLM Compiler License Agreement
- Commercial
- Allowed with conditions
- Parameters
- 7B and 13B (base and FTD variants) (about 3.9 GB at 4-bit)
Gated download. Llama acceptable use policy applies; over 700M monthly active users needs a separate Meta license.
ESM3
Partly openEvolutionaryScale · Science · Jun 25, 2024
- License
- MIT
- Commercial
- Allowed
- Parameters
- 98B flagship; also 1.4B and 7B
Only the 1.4B esm3-sm-open-v1 is downloadable; 7B and 98B are API only. Relicensed to MIT by Biohub in 2026 (was Cambrian Non-Commercial).
Chameleon (Meta AI)
Non-commercialMeta AI (FAIR) · Vision · Jun 18, 2024
- License
- Chameleon Research License
- Commercial
- Not allowed
- Parameters
- 7B and 34B
7B and 34B weights are released for research use only, after an access request.
Florence-2
PermissiveMicrosoft Research · Vision · Jun 16, 2024
- License
- MIT
- Commercial
- Allowed
- Parameters
- 232M (base); 771M (large)
Stable Diffusion 3
Partly openStability AI · Image · Jun 12, 2024
- License
- Stability AI Community License
- Commercial
- Allowed with conditions
- Parameters
- 800M to 8.1B (SD 3.5 Large)
SD3 Medium and all SD 3.5 models are downloadable; SD3 Large and Large Turbo were API only. Free commercial use under $1M annual revenue.
Weights on Hugging FaceArticle fact-checked
Qwen2
PermissiveAlibaba Cloud · Language · Jun 6, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 0.5B, 1.5B, 7B, 57B-A14B (14B active), 72B (about 0.3 GB at 4-bit)
The 0.5B, 1.5B, 7B and 57B-A14B sizes are Apache 2.0; Qwen2-72B is under the Tongyi Qianwen License (conditions on commercial use).
Weights on Hugging FaceMemory estimateArticle fact-checked
OpenVLA
Custom licenseStanford, UC Berkeley, TRI, Google DeepMind, Physical Intelligence, MIT · Robotics · Jun 2024
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 7B
Code is MIT and the Hub tag reads MIT, but the developers say the checkpoints derive from Llama 2 and fall under the Llama Community License.
Weights on Hugging FaceArticle fact-checked
Codestral
Partly openMistral AI · Language · May 29, 2024
- License
- Mistral Non-Production License
- Commercial
- Not allowed
- Parameters
- 22.2B (22B v0.1); 7.3B (Mamba); later versions not disclosed (about 12 GB at 4-bit)
Codestral 22B weights are MNPL (non-production); Codestral Mamba 7B is Apache 2.0; Codestral 25.01, 25.08 and Embed are API only.
Weights on Hugging FaceMemory estimateArticle fact-checked
DeepSeek-Prover
Custom licenseDeepSeek · Language · May 23, 2024
- License
- DeepSeek Model License
- Commercial
- Allowed with conditions
- Parameters
- 7B; 671B MoE (V2) (about 3.9 GB at 4-bit)
V1, V1.5 and V2 weights use the DeepSeek License; the V2 Hugging Face repos have no license tag and ship an MIT LICENSE file.
PaliGemma
Custom licenseGoogle · Vision · May 14, 2024
- License
- Gemma Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- 3B (PaliGemma); 3B, 10B and 28B (PaliGemma 2)
Gated on Hugging Face: the Gemma terms must be accepted before download. Covers PaliGemma and PaliGemma 2.
Article fact-checked
AlphaFold 3
Non-commercialGoogle DeepMind and Isomorphic Labs · Science · May 8, 2024
- License
- AlphaFold 3 Model Parameters Terms of Use
- Commercial
- Not allowed
Weights are for non-commercial use by non-commercial organizations only and may not be redistributed. The code is Apache 2.0 since June 2026.
Article fact-checked
DeepSeek-V2
Custom licenseDeepSeek · Language · May 2024
- License
- DeepSeek Model License
- Commercial
- Allowed with conditions
- Parameters
- 236B total; ~21B active per token (about 132 GB at 4-bit)
Code is MIT; weights use the DeepSeek License, which allows commercial use subject to use-based restrictions.
Weights on Hugging FaceMemory estimateArticle fact-checked
IBM Granite
PermissiveIBM · Language · May 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~350M to 32B MoE (H-Small, ~9B active) (about 0.2 GB at 4-bit)
The first 2023 Granite models were proprietary (watsonx only); releases since Granite Code (May 2024) are Apache 2.0.
Snowflake Arctic
PermissiveSnowflake AI Research · Language · Apr 24, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 480B total; 17B active per token (about 268 GB at 4-bit)
Phi-3
PermissiveMicrosoft · Language · Apr 23, 2024
- License
- MIT
- Commercial
- Allowed
- Parameters
- 3.8B (Mini) to 14B (Medium); Phi-3.5-MoE 42B total, 6.6B active (about 2.1 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Llama 3
Custom licenseMeta · Language · Apr 18, 2024
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 8B and 70B (base and Instruct variants) (about 4.5 GB at 4-bit)
Gated download. Acceptable use policy applies; over 700M monthly active users needs a separate Meta license.
Weights on Hugging FaceMemory estimateArticle fact-checked
Mixtral 8x22B
PermissiveMistral AI · Language · Apr 17, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~141B total; ~39B active per token (about 79 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
RecurrentGemma
Custom licenseGoogle DeepMind · Language · Apr 11, 2024
- License
- Gemma Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- 2.68B and 8.58B total (about 1.5 GB at 4-bit)
Gated on Hugging Face and Kaggle: accept the Gemma Terms of Use first. The reference code is Apache 2.0.
CodeGemma
Custom licenseGoogle · Language · Apr 9, 2024
- License
- Gemma Terms of Use
- Commercial
- Allowed with conditions
- Parameters
- 2B to 7B (about 1.1 GB at 4-bit)
Download requires accepting the Gemma Terms of Use, which include a prohibited-use policy.
Command R+
Non-commercialCohere · Language · Apr 4, 2024
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 104 billion (about 58 GB at 4-bit)
Gated on Hugging Face (contact details required); Cohere Labs Acceptable Use Policy also applies. Commercial use only through Cohere.
NVIDIA Canary
PermissiveNVIDIA · Speech and audio · Apr 2024
- License
- CC BY 4.0
- Commercial
- Allowed
- Parameters
- 182M to 2.5B across variants
The 2025 models (180M-Flash, 1B-Flash, 1B-v2, Qwen-2.5B) are CC BY 4.0; the original Canary-1B (April 2024) is CC BY-NC 4.0, non-commercial.
Reka Edge
Partly openReka AI · Language · Apr 2024
- License
- Reka Edge 2603 License (Business Source License 1.1)
- Commercial
- Allowed with conditions
- Parameters
- 7 billion (about 3.9 GB at 4-bit)
Only the 2026 redesign (reka-edge-2603) is downloadable; production use is free under $1M annual revenue. The 2024 Reka Edge was never released as weights.
Jamba
PermissiveAI21 Labs · Language · Mar 28, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 52B total; 12B active (Jamba 1.0); up to 398B total (Jamba 1.5 Large) (about 29 GB at 4-bit)
Jamba 1.0, Jamba Reasoning 3B and Jamba 2 are Apache 2.0; Jamba 1.5, 1.6 and 1.7 use the Jamba Open Model License.
DBRX
Custom licenseDatabricks (Mosaic AI) · Language · Mar 27, 2024
- License
- Databricks Open Model License
- Commercial
- Allowed with conditions
- Parameters
- 132B total; 36B active per token (about 74 GB at 4-bit)
Official Databricks Hugging Face and GitHub repos were offline in Nov 2026; mirrors keep the same license. Separate license above 700M MAU.
NVIDIA Earth-2
Custom licenseNVIDIA · Science · Mar 18, 2024
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
Varies by model: Atlas, StormScope and the first HealDA use the NVIDIA Open Model License; FourCastNet 3 and cBottle are Apache 2.0; newer ones OpenMDW 1.1.
Open-Sora
PermissiveHPC-AI Tech (Colossal-AI team) · Video · Mar 17, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 11B (v2.0)
Command R
Non-commercialCohere · Language · Mar 11, 2024
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 35B (Command R); family spans 7B to 104B (about 20 GB at 4-bit)
All open members (Command R v01 and 08-2024, R7B, R+) are CC BY-NC 4.0 and gated on Hugging Face (contact details required).
DeepSeek-VL
Custom licenseDeepSeek · Vision · Mar 11, 2024
- License
- DeepSeek Model License
- Commercial
- Allowed with conditions
- Parameters
- 1.3B and 7B
Code is MIT; weights use the DeepSeek Model License, which allows commercial use subject to use restrictions.
BitNet b1.58
PermissiveMicrosoft Research · Language · Feb 27, 2024
- License
- MIT
- Commercial
- Allowed
- Parameters
- 700M to 3B (original); 2B (BitNet b1.58 2B4T) (about 0.4 GB at 4-bit)
The official open weights are BitNet b1.58 2B4T (April 2025), under MIT.
Mistral Large
Partly openMistral AI · Language · Feb 26, 2024
- License
- Mistral Ecosystem Open-Weight License
- Commercial
- Allowed with conditions
- Parameters
- 123B (Large 2/2.1); 675B total, 41B active (Large 3) (about 69 GB at 4-bit)
Large 4: Model-as-a-Service operators need a commercial license. Large 3 is Apache 2.0; Large 2 and 2.1 are research only; Large 1 was API only.
Weights on Hugging FaceMemory estimateArticle fact-checked
Nemotron-4
Partly openNVIDIA · Language · Feb 26, 2024
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
- Parameters
- 15B; 340B (Base, Instruct, and Reward checkpoints) (about 8.4 GB at 4-bit)
The 340B Base, Instruct and Reward checkpoints are downloadable. No public Nemotron-4 15B checkpoint was found.
Gemma
PermissiveGoogle DeepMind · Language · Feb 21, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 270 million to 31 billion across generations (about 0.2 GB at 4-bit)
Gemma 4 (April 2026) is Apache 2.0; Gemma 1, 2, 3 and 3n remain under the Gemma Terms of Use, which carry a prohibited use policy.
Memory estimateArticle fact-checked
V-JEPA
Non-commercialMeta AI (FAIR) · Vision · Feb 15, 2024
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- Up to ~630M (ViT-H/16 encoder)
DeepSeekMath
Custom licenseDeepSeek · Language · Feb 5, 2024
- License
- DeepSeek Model License
- Commercial
- Allowed with conditions
- Parameters
- 7 billion (about 3.9 GB at 4-bit)
Code is MIT; weights use the DeepSeek License. The 120B-token DeepSeekMath-Corpus was not published.
Weights on Hugging FaceMemory estimateArticle fact-checked
MiniCPM
PermissiveOpenBMB · Language · Feb 1, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 0.43B (MiniCPM4-0.5B) to 13.6B (MiniCPM-MoE-8x2B) across the text models (about 0.2 GB at 4-bit)
Apache 2.0 from MiniCPM4 (June 2025). Earlier weights keep OpenBMB licenses that ask for authorization or registration for commercial use.
Memory estimateArticle fact-checked
MiniCPM-V
PermissiveOpenBMB (THUNLP and ModelBest Inc.) · Vision · Feb 1, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 1.3B to 9B, depending on version
Family. Apache 2.0 from v4.0 (latest: 4.6) and for MiniCPM-o 2.6; v1.0 to 2.5 need registration for commercial use; V 2.6 is gated.
Article fact-checked
OLMo
PermissiveAllen Institute for AI (Ai2) · Language · Feb 1, 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 1B, 7B, 13B, 32B dense; OLMoE 1B active / 7B total (about 0.6 GB at 4-bit)
Memory estimateArticle fact-checked
Aya (Expanse / Vision)
Non-commercialCohere Labs (Cohere) · Language · Feb 2024
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 8B and 32B (Aya Expanse); 13B (Aya 101) (about 4.5 GB at 4-bit)
Aya 23, Aya Expanse and Aya Vision are CC BY-NC 4.0 with Cohere's acceptable use addendum (gated click-through); the earlier Aya 101 is Apache 2.0.
Reka Flash
Partly openReka AI · Language · Feb 2024
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 21 billion (about 12 GB at 4-bit)
Reka Flash 3 and 3.1 (2025) are Apache 2.0; the original multimodal Reka Flash and its October 2024 update are API only.
AlphaGeometry
Partly openGoogle DeepMind · Science · Jan 17, 2024
- License
- CC BY 4.0
- Commercial
- Allowed
- Parameters
- Up to 3.3B (AlphaGeometry 2 language model)
AlphaGeometry 1 checkpoint and vocabulary are CC BY 4.0 (code Apache 2.0); the AlphaGeometry 2 language model weights are not released.
Article fact-checked
GLM-4
Partly openZhipu AI · Language · Jan 16, 2024
- License
- MIT
- Commercial
- Allowed
API tiers (GLM-4, Plus, Air, Flash) are closed. GLM-4-0414 weights are MIT; GLM-4-9B uses the glm-4 license (commercial use needs registration).
Article fact-checked
Minitron
Custom licenseNVIDIA · Language · 2024
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
- Parameters
- About 4B to 8B (about 2.2 GB at 4-bit)
All base checkpoints carry the NVIDIA Open Model License; the cards for the original Minitron 4B and 8B say research and development only.
InternVL
PermissiveOpenGVLab (Shanghai AI Laboratory) and academic partners · Vision · Dec 21, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 0.94B to 240.7B (28B active in InternVL3.5-241B-A28B)
InternVL3.5 is Apache 2.0. Earlier releases are MIT or follow the base LLM (Qwen license on InternVL2.5-78B and InternVL3-78B).
Article fact-checked
Mixtral
PermissiveMistral AI · Language · Dec 11, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 8x7B: 46.7B total, 12.9B active; 8x22B: 141B total, 39B active (about 3.9 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Mamba
PermissiveAlbert Gu and Tri Dao · Language · Dec 1, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 130M to 2.8B (pretrained language models) (about 0.1 GB at 4-bit)
Architecture page; the license is that of the authors' pretrained language-model checkpoints (130M to 2.8B).
Weights on Hugging FaceMemory estimateArticle fact-checked
NVIDIA Parakeet
PermissiveNVIDIA (first models with Suno.ai) · Speech and audio · Dec 2023
- License
- CC BY 4.0
- Commercial
- Allowed
- Parameters
- 110M to 1.1B
Original models and TDT 0.6B v2 and v3 are CC BY 4.0; newer variants such as parakeet-unified-en-0.6b use the NVIDIA Open Model License.
Article fact-checked
DeepSeek LLM
Custom licenseDeepSeek · Language · Nov 29, 2023
- License
- DeepSeek Model License
- Commercial
- Allowed with conditions
- Parameters
- 7B and 67B (about 3.9 GB at 4-bit)
Code is MIT; weights use the DeepSeek License, which allows commercial use subject to use-based restrictions.
Pika (video generation)
Partly openPika Labs · Video · Nov 28, 2023
- License
- MiniMax H3 Community License Agreement
- Commercial
- Allowed with conditions
Only Pika H3 Flex (LoRA adapters on MiniMax H3, Oct 2026) is open; its license excludes the EU, UK, South Korea and US. Other Pika models are hosted only.
Article fact-checked
Stable Video Diffusion
Custom licenseStability AI · Video · Nov 21, 2023
- License
- Stability AI Community License
- Commercial
- Allowed with conditions
- Parameters
- About 1.52B (1,524,623,082 in the img2vid-xt checkpoint)
Launched November 2023 as a non-commercial research preview; under the Stability AI Community License since July 2024 (free under $1M revenue).
Weights on Hugging FaceArticle fact-checked
GraphCast
PermissiveGoogle DeepMind · Science · Nov 14, 2023
- License
- CC BY 4.0
- Commercial
- Allowed
- Parameters
- 36.7M
Weights were relicensed on August 6, 2026 from CC BY-NC-SA 4.0 to CC BY 4.0, which permits commercial use; the code is Apache 2.0.
Article fact-checked
Grok
Partly openxAI · Language · Nov 4, 2023
- License
- Grok 2 Community License
- Commercial
- Allowed with conditions
- Parameters
- 314B, ~25% active (Grok-1); ~2.7T estimated (Grok-3); later models undisclosed (about 175 GB at 4-bit)
Grok-1 is Apache 2.0; Grok 2 weights (Aug 2025) allow commercial use under xAI's usage policy; Grok 3 and later are closed.
Memory estimateArticle fact-checked
DeepSeek-Coder
Custom licenseDeepSeek · Language · Nov 2, 2023
- License
- DeepSeek Model License
- Commercial
- Allowed with conditions
- Parameters
- 1.3B, 6.7B, 33B dense (V1); 16B and 236B MoE with 2.4B and 21B active (V2) (about 0.7 GB at 4-bit)
Code is MIT; V1 and V2 weights use the DeepSeek License, which allows commercial use subject to use-based restrictions.
Weights on Hugging FaceMemory estimateArticle fact-checked
Yi (language model)
Permissive01.AI · Language · Nov 2, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 6B and 34B base (family spans 1.5B to 34B) (about 3.4 GB at 4-bit)
Yi-6B and Yi-34B launched under the Yi Series Models Community License; the whole open Yi series is now Apache 2.0.
Nemotron
PermissiveNVIDIA · Language · Nov 2023
- License
- OpenMDW 1.1
- Commercial
- Allowed
- Parameters
- 8B (2023) to ~550B total MoE (Nemotron 3 Ultra, 2026) (about 4.5 GB at 4-bit)
Nemotron 3 Ultra and the 3.5 line use OpenMDW 1.1; earlier models use NVIDIA open model licenses; Nemotron-H is research only.
Memory estimateArticle fact-checked
BitNet
PermissiveMicrosoft Research · Language · Oct 17, 2023
- License
- MIT
- Commercial
- Allowed
Latent Consistency Models (LCM)
Custom licenseTsinghua University (IIIS) · Image · Oct 6, 2023
- License
- OpenRAIL
- Commercial
- Allowed with conditions
Official LCM-LoRA and LCM SDXL/SSD-1B checkpoints are OpenRAIL++-M; the first LCM_Dreamshaper_v7 checkpoint is tagged MIT.
CogVLM
Custom licenseZhipu AI and Tsinghua University (KEG) · Vision · Oct 5, 2023
- License
- CogVLM License
- Commercial
- Allowed with conditions
- Parameters
- About 17B (about 10B vision, 7B language)
Free for academic research; commercial use requires registration. CogVLM2 adds the Llama 3 license terms. Code is Apache 2.0.
SEDD (Score Entropy Discrete Diffusion)
Custom licenseStanford University (Aaron Lou, Chenlin Meng, Stefano Ermon) · Language · Oct 2023
- License
- License not stated
The code repository is MIT and links the checkpoints; the Hugging Face checkpoint repos carry no license statement of their own.
XTTS (Coqui XTTS)
Non-commercialCoqui AI · Speech and audio · Sep 30, 2023
- License
- Coqui Public Model License 1.0.0
- Commercial
- Not allowed
- Parameters
- About 750M
Coqui AI shut down in January 2024, so the separate commercial license it used to sell is no longer available.
Mistral 7B
PermissiveMistral AI · Language · Sep 27, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~7.24 billion (announced as 7.3B) (about 4.0 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Hunyuan
Partly openTencent · Language · Sep 7, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 389B total, 52B active (Hunyuan-Large); 560B total, 56B active (TurboS) (about 217 GB at 4-bit)
Hy3 and Hy4 Preview are Apache 2.0; earlier open releases use Tencent community licenses that exclude the EU, UK and South Korea; T1 and TurboS are hosted only.
Memory estimateArticle fact-checked
Nougat (model)
Non-commercialMeta AI · Vision · Aug 25, 2023
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 250M (small); 350M (base)
Code is MIT. The nougat-small repository is tagged CC BY 4.0 on Hugging Face, but Meta's README says the weights are CC-BY-NC.
Code Llama
Custom licenseMeta AI · Language · Aug 24, 2023
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 7B, 13B, 34B, 70B (base, Python, and Instruct variants) (about 3.9 GB at 4-bit)
Llama 2 Community License: acceptable use policy applies and a separate license is needed above 700M monthly users.
Weights on Hugging FaceMemory estimateArticle fact-checked
Qwen-VL
Partly openAlibaba Cloud (Qwen team) · Vision · Aug 22, 2023
- License
- Qwen License (Tongyi Qianwen)
- Commercial
- Allowed with conditions
- Parameters
- 9.6B
Qwen-VL and Qwen-VL-Chat (9.6B) are open; Qwen-VL-Plus and Qwen-VL-Max are API only. Above 100M monthly users a separate license is needed.
Article fact-checked
SeamlessM4T
Non-commercialMeta AI · Speech and audio · Aug 22, 2023
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 1.2B (Medium), 2.3B (Large and v2)
Supporting code and the w2v-BERT 2.0 encoder are MIT; SeamlessExpressive uses a separate Seamless non-commercial license.
AudioCraft
Non-commercialMeta AI (FAIR) · Speech and audio · Aug 2, 2023
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 300M to 3.3B (MusicGen); 1.5B (AudioGen)
Code and training framework are MIT; the MusicGen, AudioGen and MAGNeT weights are CC BY-NC 4.0.
BGE (BAAI General Embedding)
PermissiveBeijing Academy of Artificial Intelligence (BAAI) · Other · Aug 2, 2023
- License
- MIT
- Commercial
- Allowed
- Parameters
- 24M to 326M (v1.5 embedders)
v1.5, BGE-M3 and BGE-VL are MIT; bge-en-icl, bge-reranker-v2-m3 and bge-reasoner are Apache 2.0; the Gemma-2-based models carry Gemma terms.
SDXL (Stable Diffusion XL)
Custom licenseStability AI · Image · Jul 26, 2023
- License
- OpenRAIL
- Commercial
- Allowed with conditions
- Parameters
- ~2.6B U-Net; ~6.6B total with refiner
SDXL 1.0 base and refiner are CreativeML Open RAIL++-M; SDXL Turbo is under a non-commercial Stability AI license.
Weights on Hugging FaceArticle fact-checked
Llama 2
Custom licenseMeta · Language · Jul 18, 2023
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 7B, 13B, 70B (34B trained but withheld) (about 3.9 GB at 4-bit)
Gated download. Over 700M monthly active users needs a separate Meta license; outputs may not be used to improve other LLMs.
Memory estimateArticle fact-checked
InternLM
PermissiveShanghai AI Laboratory · Language · Jul 6, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 1.8B to 20B across generations; single 8B flagship in InternLM3 (about 1.0 GB at 4-bit)
InternLM3 and Intern-S1 are Apache 2.0; InternLM, InternLM2 and InternLM2.5 weights need an application form for commercial use.
Weights on Hugging FaceMemory estimateArticle fact-checked
MusicGen
Non-commercialMeta AI (FAIR) · Speech and audio · Jun 8, 2023
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 300M (Small), 1.5B (Medium), 3.3B (Large)
Weights are CC BY-NC 4.0; the AudioCraft code, including training code, is MIT.
Weights on Hugging FaceArticle fact-checked
Hiera
Non-commercialMeta AI (FAIR) · Vision · Jun 1, 2023
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 28M to 673M
Model weights are CC BY-NC 4.0; the code was relicensed to Apache 2.0 in March 2024.
Würstchen
PermissivePablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, Marc Aubreville · Image · Jun 1, 2023
- License
- MIT
- Commercial
- Allowed
- Parameters
- Stage A 18M; Stage B ~1B; Stage C ~0.99B
Aquila (language model)
Custom licenseBeijing Academy of Artificial Intelligence (BAAI) · Language · Jun 2023
- License
- BAAI Aquila Model License Agreement
- Commercial
- Allowed with conditions
- Parameters
- 7B, 34B, and experimental 70B (about 3.9 GB at 4-bit)
Commercial use is allowed within the license restrictions; the experimental 70B models use a separate BAAI Aquila 70B license. Code is Apache 2.0.
Phi (language model)
PermissiveMicrosoft Research · Language · Jun 2023
- License
- MIT
- Commercial
- Allowed
- Parameters
- 1.3B (Phi-1) to 15B (Phi-4-reasoning-vision) (about 0.7 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Massively Multilingual Speech (MMS)
Non-commercialMeta AI · Speech and audio · May 22, 2023
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 317M (MMS-300M), 965M (MMS-1B)
ImageBind
Non-commercialMeta AI (FAIR) · Vision · May 9, 2023
- License
- CC BY-NC-SA 4.0
- Commercial
- Not allowed
The README says CC-BY-NC 4.0, but the LICENSE file in the repository is CC BY-NC-SA 4.0. Weights are downloaded from GitHub.
Article fact-checked
StarCoder
Custom licenseBigCode project (Hugging Face and ServiceNow) · Language · May 4, 2023
- License
- OpenRAIL
- Commercial
- Allowed with conditions
- Parameters
- 15.5B (StarCoder); 3B, 7B, 15B (StarCoder2) (about 8.7 GB at 4-bit)
BigCode OpenRAIL-M with use restrictions for StarCoder and StarCoder2; the Stack training datasets are published on Hugging Face.
Weights on Hugging FaceMemory estimateArticle fact-checked
Shap-E
PermissiveOpenAI · Other · May 3, 2023
- License
- MIT
- Commercial
- Allowed
Falcon (language model)
Custom licenseTechnology Innovation Institute (TII) · Language · May 2023
- License
- TII Falcon License
- Commercial
- Allowed with conditions
- Parameters
- 90M to 180B across generations (about 0.1 GB at 4-bit)
Falcon 3 and Falcon-H1 use the TII Falcon LLM License; Falcon 7B and 40B are Apache 2.0; Falcon 180B has its own stricter TII license.
Article fact-checked
WizardLM
Custom licenseMicrosoft · Language · Apr 24, 2023
- License
- Llama Community License
- Commercial
- Allowed with conditions
License follows each base model; the latest hosted chat models are Llama 2 Community. WizardLM-2 was withdrawn from official channels.
DINOv2
PermissiveMeta AI Research (FAIR) · Vision · Apr 14, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 21M to 1.1B (ViT-S/14, ViT-B/14, ViT-L/14, ViT-g/14)
Relicensed from CC BY-NC 4.0 to Apache 2.0 on August 31, 2023. The LVD-142M training set was not released.
Weights on Hugging FaceArticle fact-checked
Qwen
Partly openAlibaba Cloud (Qwen Team) · Language · Apr 11, 2023
- License
- Qwen Community License 1.0
- Commercial
- Allowed with conditions
- Parameters
- Roughly 0.5B to over 1T across variants (about 0.3 GB at 4-bit)
Latest open release (Qwen3.8-Flash-Next). Most Qwen3 to Qwen3.8 checkpoints are Apache 2.0; Qwen3.8-Max has its own license; some tiers are hosted only.
Memory estimateArticle fact-checked
SenseNova
Partly openSenseTime · Language · Apr 10, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 2B to 27B (open-weight models); over 600B (SenseNova V6) (about 1.1 GB at 4-bit)
Flagship generations are API only. Open releases (U1, U1.5, Code2Video) are mostly Apache 2.0; some are MIT and SenseNova-Vision is CC BY-NC 4.0.
Memory estimateArticle fact-checked
Meta AI · Vision · Apr 5, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~91M (ViT-B), ~308M (ViT-L), ~636M (ViT-H)
SAM weights and code are Apache 2.0; the SA-1B training dataset is downloadable but under a separate research-only license.
Weights on Hugging FaceArticle fact-checked
DINO (computer vision)
Custom licenseMeta AI · Vision · Apr 2023
- License
- DINOv3 License
- Commercial
- Allowed with conditions
- Parameters
- 21M (ViT-S) to 6.7B (DINOv3 ViT-7B)
DINO and DINOv2 are Apache 2.0. DINOv3 allows commercial use but is gated and bars military, nuclear and other trade-controlled end uses.
University of Wisconsin-Madison, Microsoft Research and Columbia University · Vision · Apr 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 7B and 13B backbones (LLaVA-1.5); 0.5B to 110B backbones in later versions
Family. OneVision and OneVision-1.5 (latest) are Apache 2.0; Vicuna-based 1.5 and 1.6 checkpoints cite the Llama 2 Community License.
Article fact-checked
RFdiffusion
PermissiveBaker Lab, UW Institute for Protein Design · Science · Mar 30, 2023
- License
- BSD 3-Clause
- Commercial
- Allowed
Vicuna (language model)
Custom licenseLMSYS · Language · Mar 30, 2023
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- 7B, 13B, 33B (about 3.9 GB at 4-bit)
v1.5 (latest) is under the Llama 2 Community License; v0 to v1.3, including the 33B model, inherit the non-commercial LLaMA license.
Weights on Hugging FaceMemory estimateArticle fact-checked
SigLIP
PermissiveGoogle DeepMind · Vision · Mar 27, 2023
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 86M to 1B (SigLIP 2, ViT-B to ViT-g)
SigLIP and SigLIP 2 checkpoints are Apache 2.0; the WebLI training set is not public.
Article fact-checked
ChatGLM
Custom licenseZhipu AI with Tsinghua University KEG lab · Language · Mar 14, 2023
- License
- ChatGLM Model License
- Commercial
- Allowed with conditions
- Parameters
- ~6.2B (ChatGLM-6B); ~6B per later generation (about 3.5 GB at 4-bit)
Each generation (ChatGLM-6B, ChatGLM2-6B, ChatGLM3-6B) has its own Model License: free for research, commercial use after registering with Zhipu AI.
Weights on Hugging FaceMemory estimateArticle fact-checked
Alpaca (model)
Non-commercialStanford CRFM · Other · Mar 13, 2023
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 7B (13B variant documented)
Released as a weight diff; rebuilding the model needs Meta's original LLaMA 7B weights. Research use only.
Article fact-checked
CodeGeeX
Custom licenseTsinghua University (KEG) and Zhipu AI · Language · Mar 2023
- License
- CodeGeeX4 License
- Commercial
- Allowed with conditions
- Parameters
- 6B to 13B (about 3.4 GB at 4-bit)
Latest release (CodeGeeX4-ALL-9B): free for academic research, commercial use needs registration. The original 13B CodeGeeX License is non-commercial only.
LLaMA
Custom licenseMeta AI · Language · Feb 24, 2023
- License
- Llama Community License
- Commercial
- Allowed with conditions
- Parameters
- Up to 65B (LLaMA 1); ~2T total (Llama 4 Behemoth, unreleased) (about 36 GB at 4-bit)
LLaMA 1 was research-only on request; Llama 2 to 4 use versioned Community Licenses; Llama 4 Behemoth was never released.
Memory estimateArticle fact-checked
ControlNet
Custom licenseLvmin Zhang, Anyi Rao and Maneesh Agrawala (Stanford University) · Image · Feb 10, 2023
- License
- OpenRAIL
- Commercial
- Allowed with conditions
Official Stable Diffusion 1.5 checkpoints (1.0 and 1.1) are tagged OpenRAIL; the code is Apache 2.0. Third-party SDXL ControlNets carry their own licenses.
Article fact-checked
YOLOv8
Custom licenseUltralytics · Vision · Jan 10, 2023
- License
- GNU AGPL-3.0
- Commercial
- Allowed
- Parameters
- 3.2M (YOLOv8n) to 68.2M (YOLOv8x) for detection
Copyleft: GPL-3.0 at launch, AGPL-3.0 since April 14, 2023. Ultralytics also sells an Enterprise license.
SantaCoder
Custom licenseBigCode (Hugging Face and ServiceNow Research) · Language · Jan 9, 2023
- License
- OpenRAIL
- Commercial
- Allowed with conditions
- Parameters
- 1.1B (about 0.6 GB at 4-bit)
BigCode OpenRAIL-M: commercial use allowed subject to its use restrictions. Trained on The Stack v1.1, which BigCode publishes.
ConvNeXt
Non-commercialFacebook AI Research (FAIR) and UC Berkeley · Vision · Jan 2023
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 3.7M (V2-Atto) to 660M (V2-Huge); original ConvNeXt 29M to 350M
Original ConvNeXt (2022) checkpoints are MIT, commercial use allowed. ConvNeXt V2 (2023) pretrained and fine-tuned checkpoints are CC BY-NC 4.0.
Article fact-checked
I-JEPA
Non-commercialMeta AI · Vision · Jan 2023
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- About 632M (ViT-Huge)
Code and pretrained checkpoints share the repository's CC BY-NC 4.0 license.
MetaCLIP
Non-commercialMeta AI · Vision · 2023
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
MetaCLIP and MetaCLIP 2 checkpoints are CC BY-NC 4.0; the bundled open_clip code keeps its own license.
InternVideo
PermissiveOpenGVLab, Shanghai AI Laboratory; Nanjing University; SIAT, CAS · Vision · Dec 6, 2022
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- Up to 6B (InternVideo2-6B)
Most checkpoints are Apache 2.0; a few (ViCLIP, InternVideo2-Chat-8B) are MIT. Some sit behind a Hugging Face click-through gate.
Galactica (language model)
Non-commercialMeta AI (Papers with Code team) · Science · Nov 15, 2022
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- Five sizes, 125M to 120B
All five sizes are on Hugging Face under CC BY-NC 4.0; the galai inference library is Apache 2.0.
ESMFold
PermissiveMeta AI (FAIR Protein Team) · Science · Nov 1, 2022
- License
- MIT
- Commercial
- Allowed
- Parameters
- Production checkpoints use 3B ESM-2 stem; ESM-2 backbone available at 8M to 15B
Weights on Hugging FaceArticle fact-checked
EnCodec
PermissiveMeta AI (FAIR) · Speech and audio · Oct 24, 2022
- License
- MIT
- Commercial
- Allowed
Reference 24 and 48 kHz checkpoints ship with the MIT repository; the 32 kHz AudioCraft checkpoint used by MusicGen is CC BY-NC 4.0.
Article fact-checked
Stable Diffusion 1.5
Custom licenseRunway (RunwayML) · Image · Oct 20, 2022
- License
- OpenRAIL
- Commercial
- Allowed with conditions
- Parameters
- ~860M U-Net (trainable); ~123M CLIP text encoder; ~84M VAE
Runway deleted its original repository in August 2024; hash-identical community mirrors keep the CreativeML Open RAIL-M license.
BioGPT
PermissiveMicrosoft Research · Science · Oct 2022
- License
- MIT
- Commercial
- Allowed
- Parameters
- 347M (base); 1.5B (BioGPT-Large)
Whisper
PermissiveOpenAI · Speech and audio · Sep 21, 2022
- License
- MIT
- Commercial
- Allowed
- Parameters
- 39M (tiny) to 1,550M (large)
OpenAI's GitHub repository states MIT for code and weights; some Hugging Face cards (large-v3 among them) are tagged Apache 2.0.
Weights on Hugging FaceArticle fact-checked
Stable Diffusion
Custom licenseStability AI, with CompVis (LMU Munich), Runway, LAION, and EleutherAI · Image · Aug 22, 2022
- License
- Stability AI Community License
- Commercial
- Allowed with conditions
- Parameters
- 860M U-Net (SD 1.x); ~3.5B (SDXL base); 8.1B (SD 3.5 Large)
SD3 and SD 3.5 use the Stability AI Community License (free under $1M annual revenue); v1, v2 and SDXL use OpenRAIL licenses.
Article fact-checked
Atlas (language model)
Non-commercialMeta AI (FAIR) · Language · Aug 2022
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- Reader 220M to 11B, plus ~110M retriever (about 0.1 GB at 4-bit)
The repository license is CC BY-NC 4.0; no separate license is stated for the checkpoints. Wikipedia-derived corpora and indices are CC BY-SA.
GLM-130B
Non-commercialTsinghua KEG and Zhipu AI · Language · Aug 2022
- License
- GLM-130B License
- Commercial
- Not allowed
- Parameters
- 130 billion (dense) (about 73 GB at 4-bit)
Weights are sent after an application form; the license allows non-commercial research only. Code is Apache 2.0.
BLOOM (language model)
Custom licenseBigScience project (coordinated by Hugging Face) · Language · Jul 12, 2022
- License
- OpenRAIL
- Commercial
- Allowed with conditions
- Parameters
- 176 billion (smaller variants 560M to 7.1B) (about 98 GB at 4-bit)
BigScience RAIL License v1.0: commercial use allowed subject to its use restrictions. The ROOTS training corpus was not released in full.
Weights on Hugging FaceMemory estimateArticle fact-checked
No Language Left Behind (NLLB)
Non-commercialMeta AI · Language · Jul 2022
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
- Parameters
- 600M to 54.5B (about 0.3 GB at 4-bit)
All NLLB-200 checkpoints are CC BY-NC 4.0; the fairseq code is MIT.
OpenFold
PermissiveAlQuraishi lab (Columbia University) and the OpenFold Consortium · Science · Jun 2022
- License
- CC BY 4.0
- Commercial
- Allowed
Retrained OpenFold weights and OpenProteinSet are CC BY 4.0; code is Apache 2.0. The successor OpenFold3 is Apache 2.0.
Article fact-checked
OPT (Open Pre-trained Transformer)
Non-commercialMeta AI · Language · May 3, 2022
- License
- OPT-175B License Agreement
- Commercial
- Not allowed
- Parameters
- 125M to 175B (9 sizes) (about 0.1 GB at 4-bit)
Sizes 125M to 66B download directly; OPT-175B weights are granted on request to researchers. Non-commercial research use only.
GLIDE (OpenAI)
Partly openOpenAI · Image · Dec 20, 2021
- License
- MIT
- Commercial
- Allowed
- Parameters
- 3.5B (full model); ~300M released filtered model
Only the small GLIDE (filtered) model (~300M) was released; the full 3.5B model was withheld. OpenAI's model card does not recommend commercial use.
EXAONE
Partly openLG AI Research (LG Corporation) · Language · Dec 14, 2021
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 1.2B to 32B dense; 236B-A23B (K-EXAONE) and 750B-A37B (K-EXAONE 2.0) MoE (about 0.7 GB at 4-bit)
K-EXAONE 2.0 (Jul 2026) is Apache 2.0; EXAONE 3.0 to 4.5 use a non-commercial license; K-EXAONE 236B has its own license; 1.0 and 2.0 were closed.
Donut (Model)
PermissiveNAVER CLOVA · Vision · Nov 2021
- License
- MIT
- Commercial
- Allowed
- Parameters
- 143M (176M when English token embeddings are counted)
Weights on Hugging FaceArticle fact-checked
AlphaFold-Multimer
PermissiveGoogle DeepMind · Science · Oct 4, 2021
- License
- CC BY 4.0
- Commercial
- Allowed
Model parameters are CC BY 4.0; the code is Apache 2.0. Multimer v1 and v2 weights remain available but are deprecated.
RWKV
PermissiveBo Peng (BlinkDL) and the RWKV project, hosted by the Linux Foundation · Other · Aug 2021
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 169M to 14B
Architecture and model family; the official RWKV-4 to RWKV-7 checkpoints are all Apache 2.0.
Article fact-checked
Sentence Transformers project (2021 Hugging Face JAX/Flax community sprint) · Language · Aug 2021
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~22.7 million (about 0.0 GB at 4-bit)
Sentence Transformers project (2021 Hugging Face Flax/JAX community sprint) · Language · Aug 2021
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~110 million (about 0.1 GB at 4-bit)
AlphaFold
Non-commercialGoogle DeepMind · Other · Jul 15, 2021
- License
- AlphaFold 3 Model Parameters Terms of Use
- Commercial
- Not allowed
Latest release (AlphaFold 3) is non-commercial only. AlphaFold 2 and AlphaFold-Multimer parameters are CC BY 4.0 (commercial use allowed). Code is Apache 2.0.
Article fact-checked
SPLADE
Non-commercialNaver Labs Europe · Language · Jul 12, 2021
- License
- CC BY-NC-SA 4.0
- Commercial
- Not allowed
Code and checkpoints from SPLADE v2 to SPLADE-v3 are CC BY-NC-SA 4.0; naver/splade-v3 is gated on Hugging Face.
Article fact-checked
HuBERT
PermissiveMeta AI (then Facebook AI Research) · Speech and audio · Jun 14, 2021
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 95M (Base), 317M (Large), 964M (X-Large)
Weights on Hugging FaceArticle fact-checked
GPT-J
PermissiveEleutherAI · Language · Jun 9, 2021
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 6,053,381,344 (~6.05B) (about 3.4 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
GLM
Partly openTsinghua University and Zhipu AI (Z.ai) · Language · Jun 2021
- License
- MIT
- Commercial
- Allowed
- Parameters
- 110M to 744B total (40B active, GLM-5) (about 0.1 GB at 4-bit)
GLM-4.5 to 5.2 and GLM-5.3-Flash are MIT; the flagship GLM-5.3 has a custom license; GLM-4 mixed API-only and open 9B models.
Huawei PanGu
Partly openHuawei (Huawei Cloud and Noah's Ark Lab) · Language · Apr 2021
- License
- openPangu Model License Agreement 1.0
- Commercial
- Allowed with conditions
- Parameters
- 1.3B to 1.085T across the family (about 0.7 GB at 4-bit)
Only the 2025 openPangu releases (Embedded-7B, Pangu Pro MoE 72B) are open. The license bars use in the EU and requires a Powered by openPangu notice.
Switch Transformer
PermissiveGoogle (Google Brain) · Language · Jan 11, 2021
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- Up to 1.571 trillion (Switch-C, 2,048 experts) (about 878 GB at 4-bit)
Memory estimateArticle fact-checked
OpenAI · Vision · 2021
- License
- MIT
- Commercial
- Allowed
MIT is the license of OpenAI's repository, which distributes the nine checkpoints; no separate weights license is stated. Training data is not released.
Article fact-checked
Swin Transformer
PermissiveMicrosoft Research Asia · Vision · 2021
- License
- MIT
- Commercial
- Allowed
- Parameters
- 29M (Swin-T) to 197M (Swin-L); about 3B (SwinV2-G)
Architecture page; the official repository with its pretrained models is MIT. Hugging Face ports (microsoft/swin-*) are tagged Apache 2.0.
Article fact-checked
DeiT
PermissiveFacebook AI Research (FAIR) and Sorbonne University · Vision · Dec 23, 2020
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 5M to 87M
The facebookresearch/deit repository (DeiT and DeiT III code and checkpoints) is Apache 2.0.
mT5
PermissiveGoogle Research · Language · Oct 22, 2020
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 300M to 13B (five sizes) (about 0.2 GB at 4-bit)
Wav2Vec 2.0
PermissiveMeta (Facebook AI Research) · Speech and audio · Jun 20, 2020
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 95M (BASE), about 317M (LARGE)
Weights on Hugging FaceArticle fact-checked
Jukebox (OpenAI)
Non-commercialOpenAI · Speech and audio · Apr 30, 2020
- License
- OpenAI Jukebox Noncommercial Use License
- Commercial
- Not allowed
- Parameters
- About 1B (each upsampling prior) to about 5B (top-level prior)
Neither the software nor content created with it may be used commercially.
BlenderBot
PermissiveMeta AI (FAIR) · Language · Apr 29, 2020
- License
- MIT
- Commercial
- Allowed
- Parameters
- Up to 175B (BlenderBot 3) (about 98 GB at 4-bit)
Released through ParlAI (MIT); Hugging Face ports of BlenderBot 1 are tagged Apache 2.0; the BlenderBot 3 175B weights are shared by request.
ColBERT
PermissiveStanford University (Omar Khattab, Matei Zaharia) · Language · Apr 27, 2020
- License
- MIT
- Commercial
- Allowed
- Parameters
- 110M (ColBERTv2) (about 0.1 GB at 4-bit)
The official ColBERTv2 checkpoint and the reference code are MIT; third-party ColBERT models carry their own licenses.
Weights on Hugging FaceMemory estimateArticle fact-checked
Longformer
PermissiveAllen Institute for AI · Language · Apr 10, 2020
- License
- Apache 2.0
- Commercial
- Allowed
Weights on Hugging FaceArticle fact-checked
ELECTRA
PermissiveStanford University and Google Brain · Language · Mar 23, 2020
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 14M to 335M (about 0.0 GB at 4-bit)
Small, Base and Large checkpoints share the Apache 2.0 license of the google-research/electra repository.
Article fact-checked
DeBERTa
PermissiveMicrosoft Research · Language · 2020
- License
- MIT
- Commercial
- Allowed
- Parameters
- 22M backbone (v3-XSmall) to 1.5B (v2-XXLarge) (about 0.0 GB at 4-bit)
All versions (v1, v2, v3 and multilingual mDeBERTa) are MIT.
Article fact-checked
DETR
PermissiveFacebook AI Research (FAIR) · Vision · 2020
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 41M (ResNet-50) to 60M (ResNet-101)
Applies to the original DETR checkpoints; later variants such as Deformable DETR are separate releases.
Article fact-checked
OPUS-MT
PermissiveLanguage Technology Research Group, University of Helsinki · Language · 2020
- License
- CC BY 4.0
- Commercial
- Allowed
Project models are CC BY 4.0; many Hugging Face ports are tagged Apache 2.0 instead, so check the individual model card.
LayoutLM
PermissiveMicrosoft Research · Language · Dec 31, 2019
- License
- MIT
- Commercial
- Allowed
- Parameters
- ~113M (BASE) to ~343M (LARGE), v1 (about 0.1 GB at 4-bit)
MIT covers LayoutLM v1 only; LayoutLMv2, LayoutXLM and LayoutLMv3 are CC BY-NC-SA 4.0 (non-commercial).
Sentence-BERT (SBERT)
PermissiveNils Reimers and Iryna Gurevych, UKP Lab, TU Darmstadt · Language · Nov 2019
- License
- Apache 2.0
- Commercial
- Allowed
Original SBERT checkpoints are Apache 2.0; other models in the sentence-transformers namespace carry their own licenses.
Article fact-checked
XLM-RoBERTa
PermissiveFacebook AI Research (Meta AI) · Language · Nov 2019
- License
- MIT
- Commercial
- Allowed
- Parameters
- ~270M (Base); ~550M (Large); XL ~3.5B; XXL ~10.7B (about 0.2 GB at 4-bit)
T5 (language model)
PermissiveGoogle · Language · Oct 23, 2019
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 60 million to 11 billion (five sizes) (about 0.0 GB at 4-bit)
Weights on Hugging FaceArticle fact-checked
DistilGPT2
PermissiveHugging Face · Other · Oct 3, 2019
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 82M (81,912,576 learned parameters)
Weights on Hugging FaceArticle fact-checked
BART (language model)
PermissiveFacebook AI Research · Language · Oct 2019
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~140M (base); ~406M (large) (about 0.1 GB at 4-bit)
The Hugging Face facebook/bart checkpoints are tagged Apache 2.0; the original fairseq release is MIT, which fairseq applies to its pre-trained models.
Weights on Hugging FaceMemory estimateArticle fact-checked
DistilBERT
PermissiveHugging Face · Language · Oct 2019
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 66 million (about 0.0 GB at 4-bit)
Weights on Hugging FaceArticle fact-checked
ALBERT
PermissiveGoogle Research and Toyota Technological Institute at Chicago · Language · Sep 26, 2019
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 12M to 235M (about 0.0 GB at 4-bit)
Four sizes (base to xxlarge, v1 and v2) share the Apache 2.0 license of the google-research/albert repository.
Article fact-checked
RoBERTa
PermissiveFacebook AI Research; University of Washington · Language · Jul 26, 2019
- License
- MIT
- Commercial
- Allowed
- Parameters
- 125M (base), 355M (large) (about 0.1 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
XLNet
PermissiveCarnegie Mellon University and Google Brain · Language · Jun 2019
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 110M (Base), 340M (Large) (about 0.1 GB at 4-bit)
The official repository (zihangdai/xlnet), which links the released checkpoints, is Apache 2.0; the Hugging Face ports are tagged MIT.
Memory estimateArticle fact-checked
Baidu ERNIE
Partly openBaidu, Inc. · Language · Apr 19, 2019
- License
- ERNIE 5.5 Flash Base Model License Agreement v1.0
- Commercial
- Allowed with conditions
- Parameters
- Up to 424B total (47B active) in open-source ERNIE 4.5; claimed 2.4T in ERNIE 5.0 (about 237 GB at 4-bit)
Latest open release (ERNIE 5.5 Flash Base). The ERNIE 4.5 family (June 2025) is Apache 2.0; ERNIE Bot, 4.0, X1 and 5.0 are API or app only.
GPT-2
Custom licenseOpenAI · Language · Feb 14, 2019
- License
- Modified MIT
- Commercial
- Allowed
- Parameters
- 1.5B (XL); also 124M, 355M, 774M variants (about 0.8 GB at 4-bit)
Code and checkpoints are under OpenAI's Modified MIT License; WebText and the training code were not released.
Weights on Hugging FaceMemory estimateArticle fact-checked
BioBERT
PermissiveDMIS Lab, Korea University · Science · Jan 25, 2019
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 110M (Base); 340M (Large)
The dmis-lab/biobert GitHub LICENSE is Apache 2.0; the Hugging Face repositories carry no license tag.
CamemBERT
PermissiveInria (Almanach), Facebook AI Research, Sorbonne Universite · Language · 2019
- License
- MIT
- Commercial
- Allowed
- Parameters
- 110M (base); 335M (large) (about 0.1 GB at 4-bit)
EfficientNet
PermissiveGoogle Brain (Mingxing Tan, Quoc V. Le) · Vision · 2019
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 5.3M (B0) to 66M (B7)
Official TensorFlow code and checkpoints are Apache 2.0.
Article fact-checked
SciBERT
PermissiveAllen Institute for AI (AI2) · Language · 2019
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~110M (about 0.1 GB at 4-bit)
The allenai/scibert GitHub LICENSE is Apache 2.0; the Hugging Face repositories carry no license tag.
Wav2Vec
PermissiveMeta AI (formerly Facebook AI Research) · Speech and audio · 2019
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- About 34M (wav2vec 1.0) to 2B (XLS-R 2B)
Most checkpoints are Apache 2.0; the MMS weights are CC BY-NC 4.0 (non-commercial).
Article fact-checked
StyleGAN
Non-commercialNVIDIA Research · Image · Dec 12, 2018
- License
- CC BY-NC 4.0
- Commercial
- Not allowed
StyleGAN1 is CC BY-NC 4.0; StyleGAN2, StyleGAN2-ADA and StyleGAN3 use the NVIDIA Source Code License, also non-commercial.
Article fact-checked
BERT
PermissiveGoogle AI Language · Language · Oct 11, 2018
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 110M (Base); 340M (Large) (about 0.1 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Bert-base-uncased model
PermissiveGoogle AI Language · Language · Oct 2018
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~110 million (about 0.1 GB at 4-bit)
GPT
Partly openOpenAI · Language · Jun 11, 2018
- License
- Modified MIT
- Commercial
- Allowed
- Parameters
- 117M (GPT-1) to 175B (GPT-3); undisclosed from GPT-4 onward (about 0.1 GB at 4-bit)
GPT-1 (MIT) and GPT-2 (Modified MIT) weights are public; GPT-3 and later are hosted only. gpt-oss (Apache 2.0) has its own article.
Memory estimateArticle fact-checked
GPT-1
PermissiveOpenAI · Language · Jun 11, 2018
- License
- MIT
- Commercial
- Allowed
- Parameters
- 117 million (about 0.1 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
ELMo (Embeddings from Language Models)
PermissiveAllen Institute for AI and University of Washington · Language · Feb 2018
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~93.6 million (about 0.1 GB at 4-bit)
Checkpoints are hosted by AllenNLP; the official bilm-tf and AllenNLP repos are Apache 2.0, and no separate license is stated for the weights.
Article fact-checked
CycleGAN
PermissiveBerkeley AI Research (BAIR) Lab, UC Berkeley · Image · Mar 2017
- License
- BSD 3-Clause
- Commercial
- Allowed
Official repo is BSD 2-Clause and distributes the pretrained models; no separate license is stated for the weights.
ResNet
PermissiveKaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun · Vision · 2016
- License
- MIT
- Commercial
- Allowed
- Parameters
- 25.6M (ResNet-50, TorchVision count)
Architecture page; MIT is the license of the authors' model repository (ResNet-50/101/152). Ports in other libraries carry their own licenses.
Article fact-checked
YOLO (object detection)
Custom licenseJoseph Redmon and co-authors (v1 to v3); later versions by Ultralytics and other groups · Vision · Jun 2015
- License
- GNU AGPL-3.0
- Commercial
- Allowed
- Parameters
- Varies by version; 55.7M for YOLO26x
Copyleft. License of YOLO26, the latest release (Ultralytics also sells an Enterprise license); other versions are mostly GPL-3.0 or AGPL-3.0.
Article fact-checked
GoogLeNet (Inception v1)
PermissiveGoogle Research · Vision · Sep 17, 2014
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- ~5M (~6.8M counting all layers incl. auxiliary heads)
Google's official TF-Slim Inception v1 checkpoint is distributed on Kaggle Models under Apache 2.0.
Article fact-checked
VGGNet
PermissiveVisual Geometry Group, University of Oxford · Vision · Sep 4, 2014
- License
- CC BY 4.0
- Commercial
- Allowed
- Parameters
- 138.4M (VGG-16); 143.7M (VGG-19)
Article fact-checked
d3-turbo
PermissivevLLM Semantic Router team · Vision
- License
- Apache 2.0
- Commercial
- Allowed
- Parameters
- 26.1B loaded (25.6B text backbone, 461M vision encoder, 1.3M answer readout); listed as 27B by the hosted service
Ling-3.1-flash
PermissiveInclusionAI (Ant Group) · Language
- License
- MIT
- Commercial
- Allowed
- Parameters
- About 560B total, 25B active (publisher's count) (about 313 GB at 4-bit)
Ming-3.0
PermissiveinclusionAI · Speech and audio
- License
- MIT
- Commercial
- Allowed
- Parameters
- 7.9B total, about 1.3B active (tts-tiny); 9.3B total, 2.7B active (audio-tiny); 126B total, 6.5B active (audio-flash), per the model cards
Nemotron-Labs-TwoTower
Custom licenseNVIDIA · Language
- License
- NVIDIA Open Model License
- Commercial
- Allowed with conditions
- Parameters
- About 60B total (two 30B towers); about 3B active per tower per token (model card) (about 34 GB at 4-bit)
Weights on Hugging FaceMemory estimateArticle fact-checked
Octo (Robot Policy)
PermissiveOcto Model Team · Robotics
- License
- MIT
- Commercial
- Allowed
- Parameters
- 27M (Small) and 93M (Base)
MIT covers the code and checkpoints, not every contributed training dataset.
Weights on Hugging FaceArticle fact-checked
New repositories on Hugging Face
Model repositories published in the last 45 days by developers AI Wiki follows, for models in this directory. Checked daily. A new repository can be a new model, a new size or a quantized copy.
- nvidia/Kimi-K3.1-Pro-NVFP4Kimi K3.1 Pro
- MiniMaxAI/MiniMax-H3-Image-PEMiniMax H3
- tencent/Qwen3.8-27B-DSpine-b16Qwen3.8
- moonshotai/Kimi-K3.1-ProKimi K3.1 Pro
- nvidia/DeepSeek-V4.1-Flash-1017-NVFP4DeepSeek V4.1-Flash
- tencent/Hy-Memory-30B-A3B-GGUFHy Memory
- MiniMaxAI/MiniMax-M3.1-FlashMiniMax M3.1 Flash
- deepseek-ai/DeepSeek-V4.1-Flash-1017DeepSeek V4.1-Flash
- tencent/Hy-Memory-30B-A3B-FP8Hy Memory
- nvidia/Muse-Spark-Open-816B-A42B-DFlash2Muse Spark Open
- tencent/Hy-Memory-30B-A3BHy Memory
- nvidia/Nemotron-3.5-Rerank-8B-BF16Nemotron 3
- stepfun-ai/Step-5-Preview-NVFP4Step 5 Preview
- nvidia/Qwen-Image-NVFP4Qwen-Image
- Qwen/Qwen-Image-2.1-TurboQwen-Image-2.1-Turbo
- nvidia/Qwen3.8-Flash-Next-FP4-LocalQwen3.8-Flash-Next
What the labels mean
- Open weights, permissive license
- Weights are downloadable under a license that allows any use, including commercial use, modification and redistribution.
- Open weights, custom license
- Weights are downloadable under the developer's own terms. Commercial use is allowed, with conditions such as an acceptable-use policy, a revenue or user threshold, or registration.
- Open weights, non-commercial
- Weights are downloadable for research or personal use. Commercial use needs a separate agreement.
- Partly open
- Some models in this family have downloadable weights. Others are available only through an API or a product.
What this page does not say
- It does not rank models by quality. For that, see the benchmark-based comparison in Best open-source LLMs.
- A label is a summary, not legal advice. Read the license itself before you ship a product, and see the open-weight license comparison for the terms side by side.
- It lists only models with an AI Wiki article whose fact box records how the weights are released. 290 more model articles record closed weights and are left out.
- Open weights are not the whole of open source. Training code and training data are recorded on each article where the developer has published them.
Questions about open-weight models
What does open-weight mean?+
The trained parameters of the model can be downloaded, so you can run it on your own hardware, fine-tune it and keep using it if the developer changes or retires its hosted service. It says nothing by itself about the license: some open-weight models allow any use, some attach conditions, and some are for research only.
Is an open-weight model the same as an open-source model?+
Not necessarily. The Open Source AI Definition published by the Open Source Initiative also asks for the code used to train and run the model and for detailed information about the training data. This directory therefore labels models by what is verifiable from the release: whether the weights are downloadable, which license covers them and whether commercial use is allowed.
What do the license labels mean?+
Permissive: a license such as Apache 2.0 or MIT that allows any use, including commercial use, modification and redistribution. Custom license: the developer's own terms, which allow commercial use with conditions such as an acceptable-use policy or a revenue or user threshold. Non-commercial: research or personal use only. Partly open: a model family where only some members have downloadable weights.
How is the memory figure calculated?+
It is the approximate size of the weights at 4-bit quantization, from the parameter count on the model's article. It is shown for language models only and leaves out context memory and runtime overhead, so treat it as a lower bound. The VRAM calculator gives a fuller estimate. Where an article lists several sizes, the figure is for the first one listed.
Where does the data come from?+
Each row is read from the fact box of the model's AI Wiki article, which cites the developer's release, model card or license. Nothing is entered into this page by hand. If a row is wrong, the article is where to correct it: use its Edit or Report issue button.
Machine-readable: JSON feed of every row (CC BY 4.0) and an RSS feed of new releases.