Best Open-Source LLMs

As of October 30, 2026, the highest-scoring open-weight large language model is DeepSeek V4.1 Flash in its October 17 update, DeepSeek-V4.1-Flash-1017. It leads open-weight models on the Artificial Analysis Intelligence Index v5.0 with 32.5 and is released under the MIT License.^[1][2] Z.ai's GLM-5.3 is 0.8 points behind at 31.7 (both show as 32 on Artificial Analysis's charts) and is the top open model on Artificial Analysis's Coding Agent Index.^[1][3] Xiaomi's MiMo-V2.6-Pro, inclusionAI's Ling 3.1 Flash 1019, MiniMax-M3.1-Flash and StepFun's Step 5 Preview fill out the top six.^[1] All eight of the highest-scoring open models come from developers that Artificial Analysis lists as based in China.^[4] The best from elsewhere is Mistral AI's Mistral Large 4 at 26.0, whose weights came out on October 29, and the best model small enough for a single 80 GB GPU is Qwen3.8 27B at 20.1, under Apache 2.0.^[1][5][6]

One model would outrank all of them but cannot be downloaded yet. Artificial Analysis scores Moonshot AI's Kimi K3.1 Pro at 46.0 and does not list it as open-weight; Moonshot's launch post says "The full model weights will be released by Nov 4."^[7][8]

This is a ranked guide to the best open-weight LLMs by use case. A model qualifies only if its weights could be downloaded from Hugging Face on October 30, 2026. Overall standing follows the Artificial Analysis Intelligence Index v5.0, which Artificial Analysis introduced in October 2026 with a new set of nine evaluations.^[9] Its numbers sit on a different scale from the v4.1 index used in the July 2026 version of this page and from the launch-day scores that developers quote, so older figures cannot be compared with the ones below. Per-task results are Artificial Analysis's own runs unless labelled vendor-reported.

Best open-source LLM by use case: the quick verdict

Figures checked October 30, 2026. Every pick has public weights on Hugging Face. The Muse Spark Open repository is gated: downloading it means accepting Meta's license and submitting contact details for approval.^[13]

1. DeepSeek V4.1 Flash 1017: best overall

DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, calling it "the smallest model in our new architecture family, with native multimodal visual understanding", and upgraded it to the 1017 checkpoint on October 17; DeepSeek's API now serves the 1017 version.^[14][15] The 1017 card describes it as "a minor update to DeepSeek-V4.1-Flash".^[2] According to the V4.1 Flash card, the model is a multimodal mixture-of-experts design with 552 billion backbone parameters plus a separate 196-billion-parameter "Engram conditional memory" read by token lookup. It activates 8 billion parameters per token during prefill and 16 billion during decode, and its context window reaches one million tokens.^[16] The weights are MIT-licensed.^[2]

Why it wins: its v5.0 score of 32.5 is the highest of any open model.^[1] Among open models with a v5.0 score it is first on three of the nine evaluations: AA-Briefcase, which grades multi-week knowledge-work projects built from thousands of source files (Elo 1570); FrontierCode 2.0, Cognition's test of whether an agent's change to a real open-source repository is good enough for a maintainer to merge (28.5%); and Terminal-Bench-Science (12.9%).^[1][9] Its weak spot is restraint: on AA-Omniscience it answers 39.6% of questions correctly but has a 64.1% hallucination rate, against 28.0% for Muse Spark Open.^[1] DeepSeek's API charges $0.30 per million input tokens (cache miss) and $1.20 per million output tokens at peak times, and half those rates off-peak.^[15]

Best for: a default self-hosted model for mixed knowledge work, coding and analysis, where a plain MIT license matters.

Minimum hardware: the official checkpoint stores weights in FP8 with FP4 experts and totals 510.3 GB. That fits within the 640 GB of GPU memory in an 8x H100 80 GB node with about 130 GB to spare.^[2][17] DeepSeek puts the model's global KV cache footprint at 890 bytes per token, which is under 1 GB for a full million-token context.^[16]

2. GLM-5.3: best for agentic coding

Z.ai announced GLM-5.3 on August 14, 2026 and made it open-weight on August 28.^[18][19] Z.ai's model index labels it "744B-A40B" (744 billion total parameters, 40 billion active). It uses the same base model as GLM-5.2, and the launch post says its gains come from "post-training on the 743B base model".^[20][18] It accepts text only, has a 1M-token context and up to 128K output tokens, and always reasons, at low, high or max effort.^[21]

Why it wins for coding: Artificial Analysis's Coding Agent Index v2.0 tests a model inside a specific agent harness on three equally weighted benchmarks (FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA), averaging three attempts per task.^[3] GLM-5.3 running in Opencode scores 48.9, the highest of any open model. Mistral Large 4 in Mistral Vibe is effectively level at 48.3, and DeepSeek V4.1 Flash 1017 in DeepSeek's own harness scores 46.3; an earlier V4.1 Flash checkpoint scored 47.3 in OpenAI's Codex harness, a reminder that these scores belong to model-and-harness pairs.^[3] Inside the v5.0 suite, GLM-5.3 has the best open-model Terminal-Bench 4.0 result (41.9%), while DeepSeek leads on FrontierCode 2.0 (28.5% against 21.6%).^[1] Z.ai also reports 84.5 on CyberGym, a vulnerability-discovery benchmark, run in Claude Code at max effort (vendor-reported).^[22]

Best for: long-running coding and terminal agents, and defensive security work, when you can host a large model.

License: the GLM-5.3 License is MIT-style with one added condition. A licensee that runs a Model-as-a-Service business and has more than US$10 billion in revenue over 12 months "must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose".^[22] GLM-5.3-Flash uses the plain MIT license.^[23]

Minimum hardware: the FP8 checkpoint is 755.6 GB, more than the 640 GB in an 8x H100 80 GB node but less than the 1,128 GB in an 8x H200 node (141 GB per GPU). The BF16 checkpoint is 1,506.7 GB.^{[22][24][17][25]} Unsloth's 4-bit GGUF builds take 365.3 GB (UD-IQ4_XS) to 467.3 GB (UD-Q4_K_XL).^[26]

3. MiMo-V2.6-Pro: best for image, video and audio input

Xiaomi's MiMo-V2.6-Pro is a sparse mixture-of-experts model with "1.02T total / 42B activated parameters", a 1M-token context, a five-layer multi-token prediction module used for speculative decoding, and listed modalities of "Text, Image, Video, Audio".^[10] Its weights are on Hugging Face under the MIT license named in the model card; Artificial Analysis dates the release to September 21, 2026.^[10][1]

Why it wins: at 29.9 it is third among open models on v5.0, and it is the highest-scoring open model that Artificial Analysis lists as accepting text, image, video and audio. DeepSeek V4.1 Flash 1017 takes text and images, GLM-5.3 takes text only, and the next model that Artificial Analysis lists with all four input types, Muse Spark Open, scores 24.7.^[1] Xiaomi's own card for the smaller MiMo-V2.6-Flash (26.7) also lists text, image, video and audio, but Artificial Analysis lists that model as text and image only.^[41][1] Artificial Analysis measures 34.8% on Terminal-Bench 4.0, close to the 34.9 Xiaomi reports for itself.^[1][10] It also rates 52.8 on Artificial Analysis's Openness Index, ahead of DeepSeek V4.1 Flash 1017 (44.4) and GLM-5.3 (33.3).^[1]

Best for: assistants and agents that work across documents, screenshots, video and audio in one model.

Minimum hardware: the published checkpoint, stored in FP8 with MXFP4 experts, is 573.5 GB, under the 640 GB of an 8x H100 80 GB node.^[10]

4. MiniMax-M3.1-Flash: best for business workflow automation

MiniMax describes M3.1-Flash as a "native multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity" with text, image and video input, a 1M-token context and "Approximately 428B total; 23B activated per token".^[11]

Why it wins: among open models with a v5.0 score it leads three evaluations. These are AutomationBench-AA, Artificial Analysis's run of Zapier's benchmark of SaaS workflows that span several simulated business apps (72.5%); GDP.pdf, Surge AI's test of reasoning over long professional documents (20.8%); and HLE-Diamond, a 1,000-question refresh of Humanity's Last Exam (28.1%).^[1][9] Its overall score of 28.9 is fifth among open models.^[1]

Two caveats apply. Its AA-Omniscience hallucination rate is 93.7%, the second highest of any open model with a v5.0 score, so its answers need checking.^[1] Its license is also narrower than MIT: the MiniMax Community License grants rights "for non-commercial purposes", and commercial users must display "Built with MiniMax M3.1-Flash". Products that "generate more than 20 million US dollars" a year need "a separate, prior written authorization from MiniMax" (smaller ones send a one-time notice), and the license bars use for "any military purpose".^[11]

Best for: automating multi-app business workflows and question answering over long documents, with a human or verifier reviewing the output.

Minimum hardware: the official checkpoint, stored in MXFP8 with NVFP4 experts, is 252.4 GB, just under half the size of DeepSeek V4.1 Flash 1017's.^[11][2]

5. Muse Spark Open: best for factual reliability

Meta released Muse Spark Open on October 15, 2026 as "the open weights version of Muse Spark, an 816B-parameter sparse mixture-of-experts model that activates roughly 42 billion parameters per token". It takes text, image, video and audio input and holds a million tokens of context.^[27]

Why it wins: AA-Omniscience "assigns points for correct answers, subtracts points for hallucinated responses, and keeps abstentions neutral".^[9] Muse Spark Open's 25.8 on that index is the highest of any open-weight model Artificial Analysis has tested; the next is Step 5 Preview at 16.4. It answers 42.0% of questions correctly with a 28.0% hallucination rate, and it also has the best open-model result on the MMMU-Pro multimodal reasoning test (80.9%).^[1] Its overall v5.0 score is 24.7, and agentic coding is its weak area: 8.1% on Terminal-Bench 4.0, and 26.3 on the Coding Agent Index in Meta's Muse Code harness.^[1][3]

Best for: research assistants, retrieval and question answering where a wrong answer costs more than no answer.

License: the Muse Model License requires products to "prominently display" the phrase "Built with Muse Spark Open". A Model-as-a-Service provider whose group revenue "exceeds twenty (20) million US dollars" over 12 months "must notify Meta and enter into a separate written agreement with Meta"; internal use, research and commercial use that is not Model as a Service are exempt. Its "Derivatives" include any model "created, trained, modified (including fine-tuned), distilled, or otherwise improved using the Software or any outputs or results thereof".^[13]

Minimum hardware: Meta says the BF16 weights are "about 1.6 TB" and that the NVFP4 version (about 530 GB) "Serves on four Blackwell Ultra GPUs or eight H200s"; Unsloth's smallest build fits in about 300 GB of combined system RAM and VRAM.^[27] The Hugging Face repositories measure 1,644.3 GB (BF16) and 533.0 GB (NVFP4).^[13][28]

6. Mistral Large 4: best from outside China

Mistral's documentation dates the public preview of Mistral Large 4 (nicknamed "Le Chonk") to October 6, 2026, and Mistral released the weights of an improved version on October 29.^[29][5] It is "a 1 trillion-parameter natively multimodal Mixture-of-Experts model with 49 billion active parameters per token (52 billion including embeddings and output layers)", trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe. It takes text and image input, has a 1M-token context and offers low, medium and high reasoning effort.^[30]

Why it wins: at high effort it scores 26.0 on v5.0 (23.1 at medium), the highest of any open model from a developer outside China in Artificial Analysis's data; Mistral describes the score as "the highest among western open-weight models".^[1][30] On the Coding Agent Index it scores 48.3 in Mistral Vibe, 0.6 points behind GLM-5.3. That index's own Terminal-Bench 4.0 runs put it at 41.4% against GLM-5.3's 39.9%, although in the v5.0 suite GLM-5.3 leads on that test (41.9% against 24.7%).^[3][1] Mistral says the model uses "47% fewer tokens per task on average" (vendor-reported).^[3][30] Its AA-Omniscience hallucination rate of 28.2% is nearly the same as Muse Spark Open's.^[1] Mistral's API charges $0.70 per million input tokens and $2.10 per million output tokens.^[5]

Best for: teams that need a frontier-scale open model from a European developer, especially for coding agents.

License: unlike the Apache-2.0 Mistral Large 3, Large 4 uses the Mistral Ecosystem Open-Weight License, which Mistral calls "a modified version of the MIT license".^[31][30] A licensee that operates a Model-as-a-Service business "must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose"; internal use is exempt.^[30]

Minimum hardware: the main checkpoint is FP8 and takes 1,060.6 GB. That is more than an 8x H100 80 GB node holds, and it leaves about 67 GB of an 8x H200 node's 1,128 GB. A separate 610.0 GB NVFP4 checkpoint "targets Blackwell GPUs", according to Mistral.^[30][32][25]

7. Qwen3.8 27B: best on a single GPU

The Qwen team's Qwen3.8-27B is "a compact, deployment-friendly dense model: a native vision-language model that understands images and videos". It has 27 billion parameters, mixes Gated DeltaNet and Gated Attention layers, and has a context of 262,144 tokens natively, extensible to 1,000,000. Qwen published it in August 2026 under Apache 2.0.^[6][1]

Why it wins: at its highest (xhigh) effort it scores 20.1 on v5.0 (17.5 at medium, 15.7 at low), the best of any open model under 100 billion parameters. The next, Meta's Muse Glimmer 30B, scores 6.3.^[1][33] It is far behind the large models on agentic work, with 5.6% on Terminal-Bench 4.0 against 41.9% for GLM-5.3.^[1]

Best for: local and single-server deployments, including image and video understanding, where a 500 GB model is out of reach.

Minimum hardware: the BF16 weights are 55.6 GB, which fit in one 80 GB H100 with about 24 GB left for the KV cache. Qwen's official FP8 checkpoint is 30.9 GB, and Unsloth's Q4_K_M GGUF is 16.5 GB, which leaves about 7.5 GB of headroom on a GPU with 24 GB of memory.^[6][34][35]

8. Nemotron 3.5 Super VL: most open about training data and methods

NVIDIA's Nemotron 3.5 Super VL has 120 billion total and 12 billion active parameters, a 1M-token context, and interleaved Mamba-2, mixture-of-experts and attention layers with a C-RADIO vision encoder. NVIDIA gives its Hugging Face release date as October 15, 2026.^[12] It is released under OpenMDW-1.1, which grants permission "to deal in the Model Materials without restriction", asks that the license and notices be kept with copies, and "does not impose any restrictions or obligations with respect to any use, modification, or sharing of any outputs".^[12][36]

Why it wins: Artificial Analysis's Openness Index combines model availability (API access, open weights and permissive licensing) with transparency about a model's methodology and data, on a 0 to 100 scale.^[37] NVIDIA's Nemotron models score 83.3, the highest of any open model with a v5.0 score that Artificial Analysis has rated for openness (it has not yet published values for some, including Mistral Large 4 and MiniMax-M3.1-Flash), and they are the only v5.0-scored models that get any credit for disclosing pre-training data.^[1][37] Nemotron 3.5 Super VL has the highest v5.0 score of those Nemotron models (13.6), though it is well behind the leaders. Fully open research models rate higher still (88.9 for the Allen Institute for AI's Olmo 3 and the Swiss AI Initiative's Apertus, the top scores on the index) but have no v5.0 score.^[37][1]

Best for: organizations that must audit how a model was trained, or that want the closest available match to open source in the strict sense.

Minimum hardware: the BF16 weights are 248.9 GB, and NVIDIA's card lists deployment on 4x H100 or 2x GB200.^[12]

Other open models worth knowing

Ling 3.1 Flash 1019 from inclusionAI scores 29.5 on v5.0, fourth among open models, and has the best CritPt v2 hard-subset result (37.5%) of any open model with a v5.0 score. It is a text-only model with 560 billion total and 25 billion active parameters, a 1M-token context and an MIT license; its FP8 checkpoint is 573.8 GB.^[38][1]

Step 5 Preview from StepFun has 600 billion total and 27 billion active parameters, a 1M-token context and image input; StepFun said the weights would be released on October 15.^[39] It scores 28.5 on v5.0, has the second-best AA-Omniscience Index among open models (16.4), and scores 43.1 on the Coding Agent Index in Claude Code.^[1][3] The StepFun Model License 1.0 requires "Powered by StepFun AI" attribution once a product passes 100 million active users or US$20 million in revenue in a month. It also requires a separate commercial license for hardware device makers and for Model-as-a-Service businesses with more than US$20 million in revenue over 12 months.^[40] StepFun lists FP8 deployment on one 8-GPU H200 node or one 4-GPU B300 node; the checkpoint is 650.3 GB.^[40]

GLM-5.3-Flash is "the first natively multimodal model in the GLM-5 series", with 320 billion total and 18 billion active parameters on a newly trained base model that mixes sparse and linear attention. It scores 27.8 under a plain MIT license, and its FP8 checkpoint is 328.3 GB.^[23][1] Xiaomi's MiMo-V2.6-Flash (309 billion total, 15 billion active, MIT) scores 26.7 from a 177.7 GB checkpoint.^[41][1]

DeepSeek-V4-Pro-0813 is "the official release of DeepSeek-V4-Pro, superseding the preview version"; DeepSeek rolled it out on August 13, 2026.^[42][14] It keeps the preview's structure of 1.6 trillion total and 49 billion active parameters, is MIT-licensed, scores 23.1 on v5.0 and ships as an 892.7 GB checkpoint.^[42][43][1]

Kimi K3 from Moonshot AI is "an open-weight, native multimodal agentic model" with 2.8 trillion total and 104 billion active parameters and a 1M-token context.^[44] Artificial Analysis dates its release to July 2026 and now marks it deprecated in favor of Kimi K3.1 Pro. Its max-effort setting has no v5.0 score, its low-effort setting scores 17.4, and it scores 41.4 on the Coding Agent Index at max effort in Kimi Code CLI.^[45][1][3] The Kimi K3 License requires a separate agreement with Moonshot AI for Model-as-a-Service businesses with more than US$20 million in revenue over 12 months, and a "Kimi K3" display above 100 million monthly active users or US$20 million in monthly revenue. Its 1,560.9 GB checkpoint is larger than the 1,128 GB of an 8x H200 node.^[44]

Qwen3.8-2.4T-A95B is the open checkpoint behind the hosted Qwen3.8-Max, which adds "vision input & non-thinking support, 1M context length by default" and other features.^[46] Qwen's August 3 post introducing Qwen3.8-Max said the 2.4-trillion-parameter model, with 95 billion active, would have "open weights releasing next week". The Qwen3.8-Max License requires a separate license for "Model as a Service or AI Work Assistant" businesses with more than US$50 million in revenue over 12 months.^[47][46] The downloadable checkpoint has no v5.0 score; Artificial Analysis scores the hosted Qwen3.8 Max (0902) at 33.6, above every open model, but lists it as a proprietary model.^[1][33] The FP8 checkpoint is 2,496.1 GB, more than twice the memory of an 8x H200 node, and the BF16 repository is 4,892.4 GB.^[58][46][25]

Qwen3.8-Flash-Next (125 billion parameters with 6 billion active, plus a 51-billion-parameter n-gram embedding and a 4-billion-parameter multi-token prediction module) has no v5.0 score but posts the highest AA-Briefcase rating of any open model (1575, against 1570 for DeepSeek V4.1 Flash 1017).^[48][33] Its Qwen Community License 1.0 requires any "Model as a Service or AI Work Assistant business" to obtain a separate license from Qwen before commercial use, with no revenue threshold.^[48]

Three releases from US and UAE developers sit in the middle of the board: the Institute of Foundation Models' K2 Horizon 375B (17.8), Reflection AI's Beam 1.0 (17.7) and Thinking Machines Lab's Inkling (12.5).^[33][4]

Liquid AI's LFM3-1.7B, published in October 2026, is "a text model for local agents and specialized applications" with a 256K-token context that takes "about 1 GB of memory" at 4-bit precision. It has no Artificial Analysis score, and its LFM Open License v1.0 does not cover commercial use by organizations with annual revenue of US$10 million or more.^[49]

OpenAI's gpt-oss models remain under Apache 2.0. OpenAI says gpt-oss-120b (117 billion parameters, 5.1 billion active) runs "on a single 80GB GPU" and gpt-oss-20b "within 16GB of memory".^[50] Google DeepMind's Gemma 4 comes in five sizes (E2B, E4B, 12B, 26B A4B and 31B) with up to 256K tokens of context under Apache 2.0.^[51] Neither family has a v5.0 score, and Artificial Analysis lists several other recent open releases without one, so they cannot be ranked here.^[1]

Where the July 2026 picks stand

Summary comparison table

Selected open-weight LLMs with an Artificial Analysis Intelligence Index v5.0 score, ranked by that score. The table leaves out some lower-ranked scored models, including MiniMax-M3 (18.7), K2 Horizon 375B (17.8) and Beam 1.0 (17.7), which all score above Kimi K3's low-effort 17.4; Kimi K3, which Artificial Analysis marks deprecated, is included for its Coding Agent Index result.^[33] Scores are each model's highest-scoring effort setting unless marked. Coding Agent Index scores are for the model in its best-scoring harness; "n/r" means Artificial Analysis has not reported one. Checkpoint sizes are the developer's own Hugging Face repository. Where only a month is shown for the open weights, the exact day they went public could not be confirmed; dates and sizes are sourced in each model's section.^[1][3]

Which open-source LLM should you choose?

Which is best overall? DeepSeek V4.1 Flash 1017, with GLM-5.3 close enough that either is a reasonable default.^[1] Which is best for coding agents? GLM-5.3 or Mistral Large 4, which are level on the Coding Agent Index; test them in the harness you plan to use, because scores shift between harnesses.^[3] Which is best for knowledge work? DeepSeek V4.1 Flash 1017 for long document-heavy projects, and MiniMax-M3.1-Flash for multi-app SaaS automation as long as its output is verified.^[1] Which is most reliable on factual questions? Muse Spark Open, which leads AA-Omniscience by a wide margin, with Step 5 Preview next.^[1] Which handles audio and video? MiMo-V2.6-Pro or Muse Spark Open; Xiaomi also lists audio and video input for the smaller MiMo-V2.6-Flash.^[1][41] What runs on one GPU? Qwen3.8 27B; OpenAI sizes gpt-oss-120b for a single 80 GB GPU as well, but it has no v5.0 score.^[50][1] Which licenses are least restrictive? MIT (DeepSeek, MiMo, Ling, GLM-5.3-Flash), Apache 2.0 (Qwen3.8 27B, gpt-oss, Gemma 4) and OpenMDW-1.1 (Nemotron). GLM-5.3's license adds a security review only for Model-as-a-Service operators with more than US$10 billion in revenue. The Mistral, Muse, StepFun, Kimi K3 and Qwen3.8-Max licenses add conditions for Model-as-a-Service businesses or very large products, and the MiniMax Community License starts from a non-commercial grant.

How these were ranked

Overall standing follows the Artificial Analysis Intelligence Index v5.0 as published on October 30, 2026. Version 5.0 combines nine evaluations in four equally weighted categories: knowledge work (AA-Briefcase, AutomationBench-AA, GDP.pdf), coding (FrontierCode 2.0, Terminal-Bench 4.0), science (Terminal-Bench-Science, CritPt v2 hard subset) and general (AA-Omniscience, HLE-Diamond).^[9] In the leaderboard's default view of current models, 28 open-weight entries, counting each reasoning-effort setting separately, have a v5.0 score.^[33] Many well-known open models have none: gpt-oss, Gemma 4, Qwen3.5 and Llama 4 show no v5.0 value on the leaderboard, and GLM-5.2 and Kimi K2.6, which Artificial Analysis marks deprecated, show "N/A" on their own model pages.^[33][52][54]

Scores from earlier index versions are not comparable. inclusionAI's model card says Ling 3.1 Flash 1019 "scores 44 on the Artificial Analysis Intelligence Index v4.3.2"; on v5.0 the same model scores 29.5.^[38][1] The July 2026 version of this page used v4.1, so its numbers do not carry over either. Where a model is scored at several reasoning-effort settings, this page uses the highest-scoring one (max for DeepSeek and GLM-5.3, high for Mistral Large 4, xhigh for Muse Spark Open and Qwen3.8 27B) and labels any other setting it quotes.

Per-task figures are Artificial Analysis's own runs. Figures from model cards or launch posts are labelled vendor-reported. Terminal-Bench results here are version 4.0 throughout, not the 2.x or 3.0 versions that some model cards report. FrontierCode 2.0 is part of v5.0, while the Coding Agent Index uses FrontierCode v1.1.^[9][3]

Eligibility means a public Hugging Face repository on October 30, 2026. Gated repositories (Muse Spark Open, Llama 4) count, because anyone can request access under the published license. Kimi K3.1 Pro does not count, because its weights are only promised.

Minimum hardware is computed from the Hugging Face file listings of each repository (the sum of the weight files, with 1 GB = 10^9 bytes) against 80 GB per H100 and 141 GB per H200.^[17][25] These are weight sizes only: serving also needs memory for the KV cache and activations, so a checkpoint that only just fits leaves little room for long contexts. This page makes no speed claims.

"Open source" in the title follows common usage. The Open Source Initiative's Open Source AI Definition 1.0 requires "Data Information", "Code" and "Parameters", each under OSI-approved terms, including "the complete source code used to train and run the system".^[57] Every model this page ranks publishes its parameters, but among models with a v5.0 score only NVIDIA's Nemotron family gets any credit from Artificial Analysis for disclosing pre-training data, and several of the leaders use licenses with commercial conditions.^[1] "Open-weight" is the accurate term for most of them.

Continue reading

• Artificial AnalysisArtificial Analysis is an independent benchmarking and analytics platform that evaluates artificial intelligence models and API providers across intelligence, speed, price, and latency, and it is best known…Mentioned 36 times above
• GLM-5.3GLM-5.3 is a text-input and text-output large language model from Z.ai, available through hosted services and downloadable checkpoints.Mentioned 25 times above
• MMIT LicenseThe MIT License is a short, permissive software license. It grants broad permission to use, copy, modify, merge, publish, distribute, sublicense, and sell covered software, subject to one express preservation…Mentioned 21 times above
• Qwen3.8Qwen3.8 is the name Qwen uses for a model generation that includes hosted services and downloadable checkpoints.Mentioned 16 times above
• DeepSeek V4.1-FlashDeepSeek V4.1-Flash is an open-weight multimodal mixture-of-experts model released by DeepSeek on September 10, 2026. It accepts text and images and generates text.Mentioned 12 times above
• Mistral Large 4Mistral Large 4 is a general-purpose multimodal model developed by Mistral AI. It entered public preview on October 6, 2026, with a mixture-of-experts architecture and API access.Mentioned 11 times above

More to read

• Z.ai
• Xiaomi MiMo-V2.6
• MiniMax M3.1 Flash
• Step 5 Preview
• Ling-3.1-flash
• Mistral AI

Topics:AI ModelsLarge Language ModelsOpen Source AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53 ^54Artificial Analysis, "Comparison of Open Source AI Models across Intelligence, Performance, Price, Context Window, and more" (Intelligence Index v5.0 scores, per-evaluation results and Openness Index values; accessed October 30, 2026). artificialanalysis.ai/...open-source
2. ^1 ^2 ^3 ^4 ^5 ^6DeepSeek, "DeepSeek-V4.1-Flash-1017" model card, config.json and LICENSE, Hugging Face (accessed October 30, 2026). huggingface.co/...DeepSeek-V4.1-Flash-1017
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Artificial Analysis, "AI Coding Agent Benchmarks & Leaderboard" (Coding Agent Index v2.0; accessed October 30, 2026). artificialanalysis.ai/...coding-agents
4. ^1 ^2Artificial Analysis, "Comparison of AI Models across Intelligence, Performance, and Price" (model creators and countries; accessed October 30, 2026). artificialanalysis.ai/models
5. ^1 ^2 ^3Mistral AI, "Mistral Large 4 (Le Chonk) is now open weight," October 29, 2026. mistral.ai/...mistral-large-4-oss
6. ^1 ^2 ^3 ^4Qwen, "Qwen3.8-27B" model card and LICENSE (Apache License 2.0), Hugging Face (accessed October 30, 2026). huggingface.co/...Qwen3.8-27B
7. ^Artificial Analysis, "Kimi K3.1 Pro - Intelligence, Performance & Price Analysis" (accessed October 30, 2026). artificialanalysis.ai/...kimi-k3-1-pro
8. ^Moonshot AI, "Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence" (accessed October 30, 2026). kimi.ai/...kimi-k3-1-pro
9. ^1 ^2 ^3 ^4 ^5 ^6Artificial Analysis, "Intelligence Benchmarking" (Intelligence Index v5.0 methodology; accessed October 30, 2026). artificialanalysis.ai/...intelligence-benchmarking
10. ^1 ^2 ^3 ^4 ^5Xiaomi MiMo, "MiMo-V2.6-Pro-RL" model card, Hugging Face (accessed October 30, 2026). huggingface.co/...MiMo-V2.6-Pro-RL
11. ^1 ^2 ^3 ^4MiniMax, "MiniMax-M3.1-Flash" model card, config.json and LICENSE (MiniMax Community License), Hugging Face (accessed October 30, 2026). huggingface.co/...MiniMax-M3.1-Flash
12. ^1 ^2 ^3 ^4NVIDIA, "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16" model card, Hugging Face (accessed October 30, 2026). huggingface.co/...tron-3.5-Super-VL-120B-A12B-BF16
13. ^1 ^2 ^3 ^4Meta, "Muse-Spark-Open-816B-A42B" model card and Muse Model License (gated repository), Hugging Face (accessed October 30, 2026). huggingface.co/...Muse-Spark-Open-816B-A42B
14. ^1 ^2DeepSeek API Docs, "Change Log" (entries of August 13, September 10 and October 17, 2026; accessed October 30, 2026). api-docs.deepseek.com/updates
15. ^1 ^2DeepSeek API Docs, "Models & Pricing" (accessed October 30, 2026). api-docs.deepseek.com/...pricing
16. ^1 ^2DeepSeek, "DeepSeek-V4.1-Flash" model card, Hugging Face (accessed October 30, 2026). huggingface.co/...DeepSeek-V4.1-Flash
17. ^1 ^2 ^3NVIDIA, "H100 GPU" product page (GPU memory specifications; accessed October 30, 2026). nvidia.com/...h100
18. ^1 ^2Z.ai (@Zai_org), "Introducing GLM-5.3: Built to Code. Ready for Cyber Defense," X, August 14, 2026. x.com/...2088132965922476159
19. ^Z.ai (@Zai_org), "GLM-5.3 is now open-weight," X, August 28, 2026. x.com/...2093354097122455713
20. ^1 ^2Z.ai (zai-org), "GLM-5" GitHub repository README, model download table (accessed October 30, 2026). github.com/...GLM-5
21. ^Z.ai Developer Document, "GLM-5.3 - Overview" (accessed October 30, 2026). docs.z.ai/...glm-5.3
22. ^1 ^2 ^3Z.ai (zai-org), "GLM-5.3" model card and LICENSE (GLM-5.3 License), Hugging Face (accessed October 30, 2026). huggingface.co/...GLM-5.3
23. ^1 ^2Z.ai (zai-org), "GLM-5.3-Flash" model card and LICENSE (MIT License), Hugging Face (accessed October 30, 2026). huggingface.co/...GLM-5.3-Flash
24. ^Z.ai (zai-org), "GLM-5.3-BF16" model repository, Hugging Face (accessed October 30, 2026). huggingface.co/...GLM-5.3-BF16
25. ^1 ^2 ^3 ^4NVIDIA, "H200 GPU" product page ("141 gigabytes (GB) of HBM3e"; accessed October 30, 2026). nvidia.com/...h200
26. ^Unsloth, "GLM-5.3-GGUF" model repository, Hugging Face (accessed October 30, 2026). huggingface.co/...GLM-5.3-GGUF
27. ^1 ^2Meta, "Build with Muse Spark Open," October 15, 2026. developer.meta.com/...build-with-muse-spark-open
28. ^Meta, "Muse-Spark-Open-816B-A42B-NVFP4" model repository (gated), Hugging Face (accessed October 30, 2026). huggingface.co/...Muse-Spark-Open-816B-A42B-NVFP4
29. ^Mistral Docs, "Mistral Large 4" model page (accessed October 30, 2026). docs.mistral.ai/...mistral-large-4-0
30. ^1 ^2 ^3 ^4 ^5 ^6 ^7Mistral AI, "Mistral-Large-4-1T-A52B" model card and LICENSE (Mistral Ecosystem Open-Weight License), Hugging Face (accessed October 30, 2026). huggingface.co/...Mistral-Large-4-1T-A52B
31. ^Mistral AI, "Mistral-Large-3-675B-Instruct-2512" model card, Hugging Face (accessed October 30, 2026). huggingface.co/...Mistral-Large-3-675B-Instruct-2512
32. ^Mistral AI, "Mistral-Large-4-1T-A52B-NVFP4" model repository, Hugging Face (accessed October 30, 2026). huggingface.co/...Mistral-Large-4-1T-A52B-NVFP4
33. ^1 ^2 ^3 ^4 ^5 ^6 ^7Artificial Analysis, "LLM Leaderboard" (accessed October 30, 2026). artificialanalysis.ai/...models
34. ^Qwen, "Qwen3.8-27B-FP8" model repository, Hugging Face (accessed October 30, 2026). huggingface.co/...Qwen3.8-27B-FP8
35. ^Unsloth, "Qwen3.8-27B-GGUF" model repository, Hugging Face (accessed October 30, 2026). huggingface.co/...Qwen3.8-27B-GGUF
36. ^OpenMDW, "OpenMDW-1.1" license text (accessed October 30, 2026). openmdw.ai/...1-1
37. ^1 ^2 ^3Artificial Analysis, "Artificial Analysis Openness Index" (accessed October 30, 2026). artificialanalysis.ai/...l-analysis-openness-index
38. ^1 ^2inclusionAI, "Ling-3.1-flash-1019" model card, Hugging Face (accessed October 30, 2026). huggingface.co/...Ling-3.1-flash-1019
39. ^StepFun, "Step 5 Preview: Advancing the Pareto Frontier" (accessed October 30, 2026). stepfun.com/step-5-preview
40. ^1 ^2StepFun, "Step-5-Preview" model card and LICENSE (StepFun Model License 1.0), Hugging Face (accessed October 30, 2026). huggingface.co/...Step-5-Preview
41. ^1 ^2 ^3Xiaomi MiMo, "MiMo-V2.6-Flash-RL" model card, Hugging Face (accessed October 30, 2026). huggingface.co/...MiMo-V2.6-Flash-RL
42. ^1 ^2 ^3DeepSeek, "DeepSeek-V4-Pro-0813" model card and LICENSE, Hugging Face (accessed October 30, 2026). huggingface.co/...DeepSeek-V4-Pro-0813
43. ^DeepSeek, "DeepSeek-V4-Pro" (preview) model card, Hugging Face (accessed October 30, 2026). huggingface.co/...DeepSeek-V4-Pro
44. ^1 ^2Moonshot AI, "Kimi-K3" model card and LICENSE (Kimi K3 License), Hugging Face (accessed October 30, 2026). huggingface.co/...Kimi-K3
45. ^Artificial Analysis, "Kimi K3 (max) - Intelligence, Performance & Price Analysis" (accessed October 30, 2026). artificialanalysis.ai/...kimi-k3
46. ^1 ^2 ^3Qwen, "Qwen3.8-2.4T-A95B" model card and LICENSE (Qwen3.8-Max License), Hugging Face (accessed October 30, 2026). huggingface.co/...Qwen3.8-2.4T-A95B
47. ^Qwen Team, "Qwen3.8-Max: A New Bar for Coding and Cowork," August 3, 2026. qwen.ai/blog
48. ^1 ^2Qwen, "Qwen3.8-Flash-Next" model card and LICENSE (Qwen Community License 1.0), Hugging Face (accessed October 30, 2026). huggingface.co/...Qwen3.8-Flash-Next
49. ^Liquid AI, "LFM3-1.7B" model card and LICENSE (LFM Open License v1.0), Hugging Face (accessed October 30, 2026). huggingface.co/...LFM3-1.7B
50. ^1 ^2OpenAI, "gpt-oss-120b" model card, Hugging Face (accessed October 30, 2026). huggingface.co/...gpt-oss-120b
51. ^1 ^2Google DeepMind, "gemma-4-31B-it" model card, Hugging Face (accessed October 30, 2026). huggingface.co/...gemma-4-31B-it
52. ^1 ^2Artificial Analysis, "GLM-5.2 (max) - Intelligence, Performance & Price Analysis" (accessed October 30, 2026). artificialanalysis.ai/...glm-5-2
53. ^Z.ai, "GLM-5.2: Built for Long-Horizon Tasks," June 16, 2026. z.ai/...glm-5.2
54. ^1 ^2Artificial Analysis, "Kimi K2.6" model page (accessed October 30, 2026). artificialanalysis.ai/...kimi-k2-6
55. ^Artificial Analysis, "Mistral Large 3" model page (accessed October 30, 2026). artificialanalysis.ai/...mistral-large-3
56. ^Meta, "Llama-4-Maverick-17B-128E-Instruct" model page and Llama 4 Community License (gated repository), Hugging Face (accessed October 30, 2026). huggingface.co/...Llama-4-Maverick-17B-128E-Instruct
57. ^Open Source Initiative, "The Open Source AI Definition 1.0" (read via the Internet Archive copy of October 29, 2026, web.archive.org/...open-source-ai-definition). opensource.org/...open-source-ai-definition
58. ^Qwen, "Qwen3.8-2.4T-A95B-FP8" model repository, Hugging Face (accessed October 30, 2026). huggingface.co/...Qwen3.8-2.4T-A95B-FP8

Show all 58 references