Proprietary vs. Open Source Large Language Models (LLMs)
See also: LLMs, GPT-4, Llama, Apache 2.0 License
The split between proprietary and open-weight large language models (LLMs), the latter often loosely called "open-source", is one of the defining structural choices in modern artificial intelligence. On one side are closed, API-only systems built and operated by companies like OpenAI, Anthropic, and Google DeepMind. On the other are downloadable models with publicly released weights, often produced by companies like Meta AI, Mistral AI, DeepSeek, and Alibaba's Qwen team, as well as nonprofit groups like Allen AI and EleutherAI. The two camps differ in licensing, deployment, cost structure, customization, transparency, and the kind of risk a user takes on. This article covers the differences in detail, surveys the major models in each category as of November 2026, and walks through the practical tradeoffs that engineering and product teams weigh when choosing between them.
what "proprietary" and "open source" actually mean
A proprietary LLM is a model whose weights are held privately by the developer. The model is normally accessed through a paid API, a chat product, or a partner cloud (for example, Anthropic's Claude on Amazon Bedrock). The customer does not download the network parameters and cannot run the model on their own hardware. Use is governed by the provider's terms of service, which usually include rate limits, content policies, and per-token pricing.
An open-weight LLM (often loosely called an open-source LLM) is a model whose trained parameters are publicly downloadable, normally from a hub like Hugging Face. The user can run the model on their own infrastructure, fine-tune it, redistribute it (subject to its license), and inspect its size and architecture. Some open releases also include the training code, training data, intermediate checkpoints, or the alignment recipe; the most thorough releases publish all of these together.
The boundary is not always clean. A model can be:
- Open weights only: weights are downloadable, but the training code, data mix, and recipe are private. Most popular "open" models fall here, including all current Llama and Qwen releases.
- Source-available: weights and code are published but the license restricts who can use the model or for what purpose. Llama is the canonical example.
- Fully open: weights, training code, training data, and detailed methodology are released under a permissive license. This is much rarer; Allen AI's OLMo family and EleutherAI's Pythia are the clearest examples.
- Proprietary: nothing of substance is released except an API.
the terminology problem
The phrase "open source" is used loosely in the LLM space, in ways that would not pass muster for traditional software. The Open Source Initiative (OSI) addressed this on October 28, 2024 with version 1.0 of the Open Source AI Definition.[1] To qualify as open-source AI under the definition, a system must give users freedom to use, study, modify, and share it for any purpose, and must publish three things together: the model parameters (weights), the source code used to train and run the system, and sufficiently detailed information about the training data so that a skilled person could build a substantially equivalent system. Crucially, the definition does not require releasing the raw training data itself, partly because training corpora often include copyrighted, private, or personally identifiable material. It does require a complete description of all the data used, a listing of the publicly available training data and where to obtain it, and a listing of data obtainable from third parties.[1]
By this strict bar, very few "open" LLMs qualify. Llama, Qwen, Mistral's open releases, and DeepSeek typically publish the weights but withhold the training data and detailed data documentation. Allen AI's OLMo 2 and OLMo 3, along with EleutherAI's Pythia, are among the few well-known model families that publish the weights, code, training data, training recipes, and intermediate checkpoints.[15]
A more accurate label for most "open" LLMs is open-weights or source-available. The looser usage continues, and it is disputed: critics call it openwashing, while Meta responded to the OSI definition by saying that "there is no single open source AI definition".[41]
license taxonomy
LLM licenses fall into a small number of recognizable buckets. The license matters: it determines whether you can use the model commercially, whether you can fine-tune it on your own data, whether you can use it to train a competing model, and whether your downstream users have any obligations. The table below summarizes the most common license categories.
| License type | Examples | Commercial use | Modification and fine-tuning | Notable restrictions |
|---|---|---|---|---|
| Permissive licenses on the weights (Apache 2.0, MIT, BSD, OpenMDW) | Mistral 7B, Mixtral 8x7B, Falcon, Pythia, OLMo, gpt-oss, DeepSeek-V3.1 weights; from 2026, Gemma 4 and Muse Glimmer (Apache 2.0), DeepSeek V4 (MIT), and NVIDIA's Nemotron 3 Ultra and Nemotron 3.5 models (OpenMDW-1.1)[31][82][49][79][80] | Allowed | Allowed | Patent grant (Apache 2.0) and attribution requirements |
| Restricted source-available (Llama Community License, Gemma Terms of Use, NVIDIA Open Model License) | Llama 2, Llama 3, Llama 3.1, Llama 3.3, Llama 4, Gemma releases through Gemma 3 (Gemma 4 is Apache 2.0[31]), NVIDIA's Nemotron-4 340B[91] | Allowed with conditions | Allowed | 700 million MAU threshold on the release date (Llama); ban on using outputs to improve other LLMs (Llama 2 and Llama 3 only, removed from Llama 3.1 onward); Acceptable Use Policy (Llama) and Prohibited Use Policy (Gemma)[2][3]; rights end automatically if a user circumvents a model's safety guardrails without a substantially similar replacement (NVIDIA)[92] |
| Responsible AI (OpenRAIL family) | BLOOM, several Stability AI releases, the original DeepSeek-V3 weights (DeepSeek License Agreement 1.0, a custom license modeled on OpenRAIL)[45] | Allowed | Allowed | Behavioral restrictions on harmful uses, military uses, etc. |
| Research / non-commercial | Llama 1 (original release), OPT, several academic releases; Creative Commons NonCommercial terms such as CC BY-NC 4.0 on Cohere's Command A[93] | Not allowed | Allowed for research | Output is for research only; cannot be deployed in products |
| Model-as-a-Service conditions (2026) | Qwen3.8-Max License, GLM-5.3 License, Kimi K3 and Kimi K3.1-Pro licenses, Mistral Ecosystem Open-Weight License | Allowed, but some businesses that host the model for others need a separate license or a review | Allowed | MIT-style grant plus a condition on companies that run a Model-as-a-Service business, usually above a revenue threshold[56][61][64][66][77] |
| Custom restricted | Various smaller releases | Varies | Varies | Read each license individually |
Meta's Llama community licenses, one per model version, are the best-known example of a non-permissive license for open-weight models. It allows broad commercial use with conditions. A licensee whose products had more than 700 million monthly active users in the calendar month before the version's release date must request a license from Meta, which Meta may grant at its sole discretion, and all use must follow Meta's Acceptable Use Policy. The Llama 2 and Llama 3 licenses add a competitor restriction barring use of the model or its outputs to improve any other large language model (other than a Llama derivative).[2] Meta removed that restriction from the Llama 3.1 license in July 2024, and the Llama 3.3 and Llama 4 licenses do not contain it.[3][30] The 700-million-user threshold reaches only a handful of hyperscalers. Where the competitor restriction applies (Llama 2 and Llama 3), a licensee may not use the model's outputs to distill into a non-Llama base model; from Llama 3.1 onward the license allows it.[3][30]
In 2026 several developers of the largest open-weight models adopted a newer kind of license: short texts that start from the MIT License's permission grant and add a condition aimed at companies that run a Model-as-a-Service business, selling hosted access to language models. The Qwen3.8-Max License requires a licensee running a "Model as a Service or AI Work Assistant business" with group revenue above US$50 million in any 12 months to obtain a separate license from Qwen before commercial use.[56] The Kimi K3.1-Pro License sets the same kind of requirement at US$20 million,[66] the GLM-5.3 License requires a Model-as-a-Service operator with more than US$10 billion in revenue to "pass Z.AI's security review",[61] and the Mistral Ecosystem Open-Weight License requires a separate commercial license from any licensee whose group operates a Model as a Service business, whatever its size.[77] All four define the term as giving a third party access to inference or fine-tuning ("post-training" in Mistral's text) in a way that lets it "exercise meaningful control over the inputs, parameters, or training data".[56][66][61][77]
The G7 has since given these gradations names. On May 29, 2026 the G7 digital and technology ministers approved a non-binding "Vision on AI Openness Opportunities and Shared Language", which states that "The openness of an AI is not binary" and sets out four tiers.[94] In it, "Open Weights AI" is AI released free of charge with its weights and deployment code under an Open Source license, while "Weights Available AI" is AI "released free of charge along with its weights and deployment code under a license which includes restrictions on how they can be used, such as commercial, geographical, or use-case restrictions".[94]
major proprietary LLMs
The proprietary tier is dominated by a small number of well-funded labs. As of November 17, 2026, the best-known closed-weight model families and their flagship releases are:
| Model family | Provider | Flagship release (date) | Modalities | Access |
|---|---|---|---|---|
| GPT-4, GPT-4o, GPT-4.1, o-series, GPT-5 | OpenAI | GPT-6 Astra (Sep 3, 2026)[23]; earlier GPT-5 (Aug 7, 2025)[10] and GPT-4o (May 13, 2024)[12] | Text, image, audio (GPT-4o); reasoning (o-series) | API; ChatGPT; Microsoft Copilot |
| Claude 3, 3.5, 4, 4.5, 4.6, Opus 4.7 | Anthropic | Claude Fable 5.2 (Nov 9, 2026)[24]; earlier Claude Opus 5.5 (Sep 22, 2026)[25] and Claude Opus 4.7 (Apr 16, 2026)[13] | Text, vision | API; Claude.ai; AWS Bedrock; Google Vertex AI; Microsoft Foundry |
| Gemini 1.5, 2.0, 2.5, 3 | Google DeepMind | Gemini 4 Carbon (announced Oct 28, 2026)[26]; earlier Gemini 3 Pro (Nov 18, 2025)[9] | Text, image, audio, video | Gemini app; AI Studio; Vertex AI |
| Grok | xAI | Grok 4.7 (Sep 21, 2026)[27]; earlier Grok 4 (Jul 9, 2025) | Text, vision | X (Twitter); xAI API |
Proprietary models tend to lead the published frontier benchmarks, particularly on hard reasoning and agentic-coding tasks. In early April 2026, Anthropic's Claude Opus 4.6 Thinking topped the LMArena (formerly Chatbot Arena) text leaderboard, with Gemini 3.1 Pro Preview third and GPT-5.4 High sixth.[20][42]
major open-weight LLMs
Open-weight releases multiplied rapidly after the Llama 1 leak in March 2023. The table below covers the most widely used open-weight families, followed by the major open-weight language model releases of 2026 through November; the list of open-weight models is more complete.
| Model family | Developer | Notable releases | License | Architecture notes |
|---|---|---|---|---|
| Llama | Meta AI | Llama 1 (Feb 2023), Llama 2 (Jul 2023), Llama 3 (Apr 2024), Llama 3.1, Llama 3.3 (Dec 6, 2024, 70B)[22], Llama 4 Scout/Maverick (Apr 5, 2025)[4] | Llama Community License | Llama 4 introduced Mixture-of-Experts: Scout has 17B active / 109B total / 16 experts; Maverick has 17B active / 400B total / 128 experts[21] |
| Mistral | Mistral AI | Mistral 7B, Mixtral 8x7B, Mixtral 8x22B, Mistral Large 2, Mistral 3 family and Mistral Large 3 (Dec 2025, 41B active / 675B total)[14] | Apache 2.0 (most open releases); Mistral Research License (some larger models) | Sliding window attention; sparse MoE in larger releases |
| DeepSeek | DeepSeek (China) | DeepSeek-V2, DeepSeek-V3 (Dec 26, 2024, 671B / 37B active)[6], DeepSeek-R1 (Jan 20, 2025, MIT-licensed reasoning model)[5], V3.1 | MIT (R1, V3.1 weights); custom OpenRAIL-derived (original V3 weights) | MoE with multi-head latent attention; R1 trained with reinforcement learning for chain-of-thought reasoning |
| Qwen | Alibaba | Qwen 2, Qwen 2.5, Qwen 3 (Apr 29, 2025)[7] | Apache 2.0 for most sizes; Qwen License (Qwen2.5-72B) and Qwen Research License (Qwen2.5-3B) | Dense (0.6B, 1.7B, 4B, 8B, 14B, 32B) and MoE (30B-A3B, 235B-A22B); 36T training tokens; 119 languages[7] |
| Gemma | Gemma, Gemma 2, Gemma 3 (Mar 12, 2025)[8], Gemma 3 270M, Gemma 4 (Apr 2, 2026)[31] | Gemma Terms of Use through Gemma 3 (custom terms with a prohibited-use policy); Apache 2.0 from Gemma 4[31] | Gemma 3 sizes are 1B, 4B, 12B, 27B; 4B and up are vision-language with a 128k context (32k for 1B and 270M)[8]; Gemma 4 launched as E2B, E4B, 26B MoE and 31B[31] | |
| Falcon | Technology Innovation Institute (UAE) | Falcon-7B, Falcon-40B, Falcon-180B | Apache 2.0 (most variants) | Early high-profile permissive open release in 2023 |
| OLMo | Allen Institute for AI | OLMo, OLMo 2 (7B, 13B, 32B), OLMo 3 | Apache 2.0 | Fully open: weights, training data, code, recipes, and intermediate checkpoints all released[15] |
| Pythia / GPT-NeoX | EleutherAI | Pythia (70M to 12B suite) | Apache 2.0 | Designed primarily for interpretability and scaling research; full data and checkpoints |
| gpt-oss | OpenAI | gpt-oss-120b, gpt-oss-20b (Aug 5, 2025)[11] | Apache 2.0 | OpenAI's first open-weight LLMs since GPT-2; 120B variant runs on a single 80 GB GPU |
| Qwen3.5 to Qwen3.8 | Alibaba | Qwen3.5-397B-A17B (Feb 16, 2026), Qwen3.6-35B-A3B (Apr 16, 2026), Qwen3.8-2.4T-A95B (Aug 12, 2026), Qwen3.8-27B (Aug 14, 2026)[51] | Apache 2.0 (Qwen3.5, Qwen3.6, Qwen3.8-27B)[52][53][54]; Qwen3.8-Max License (Qwen3.8-2.4T-A95B)[56] | Qwen3.5 flagship: 397B total / 17B active;[52] Qwen3.8-2.4T-A95B: 2.4T total / 95B active, text only, always runs in thinking mode, the first Qwen-Max-class model released openly[55][51] |
| GLM-5 series | Zhipu AI (Z.ai) | GLM-5 (Feb 2026), GLM-5.1 (Apr 2026), GLM-5.2 (Jun 2026), GLM-5.3 (Aug 2026)[57] | MIT (GLM-5 to GLM-5.2)[58][59][60]; GLM-5.3 License (security review for Model-as-a-Service operators above US$10 billion in revenue)[61] | GLM-5 scaled from GLM-4.5's 355B parameters (32B active) to 744B (40B active), with DeepSeek Sparse Attention[58] |
| Mistral (2026) | Mistral AI | Mistral Small 4 (Mar 16, 2026)[71], Mistral Medium 3.5[72], Mistral Large 4 (preview Oct 6, 2026; weights Oct 29, 2026)[74][75] | Apache 2.0 (Small 4)[71]; Modified MIT that excludes companies with more than US$20 million in monthly revenue (Medium 3.5)[73]; Mistral Ecosystem Open-Weight License (Large 4)[77] | Small 4: 119B total MoE;[71] Medium 3.5: 128B dense;[72] Large 4: about 1T total, 49B active per token (52B counting embeddings and output layers), natively multimodal[76] |
| DeepSeek V4 | DeepSeek | V4-Pro and V4-Flash preview (Apr 24, 2026)[47], DeepSeek V4.1-Flash (Sep 10, 2026)[48] | MIT[49][50] | V4-Pro: 1.6T total / 49B active; V4-Flash: 284B / 13B; both with a 1M-token context;[47] V4.1-Flash: causal encoder-decoder MoE with 552B backbone parameters (plus a 196B Engram memory) and image input, activating 8B parameters per token in prefill and 16B in decode[50] |
| Nemotron 3 and Nemotron 3.5 | NVIDIA | Nemotron 3 Super (Mar 11, 2026)[78], Nemotron 3 Ultra (Jun 4, 2026)[79], Nemotron 3.5 Lightning (Aug 11, 2026)[80], Nemotron 3.5 Super VL (Oct 15, 2026)[81] | NVIDIA Nemotron Open Model License (Nemotron 3 Super)[78]; OpenMDW-1.1 (Nemotron 3 Ultra, Nemotron 3.5)[79][80][81] | Hybrid Mamba-2, attention and mixture-of-experts layers: Super 120B total / 12B active, Ultra 550B / 55B, 3.5 Lightning 30B / 3B, 3.5 Super VL 120B / 12B with image and video input[78][79][80][81] |
| Hy3 | Tencent | Hy3 preview (Apr 23, 2026), Hy3 (Jul 6, 2026)[87] | Apache 2.0 (Hy3)[88] | MoE, 295B total / 21B active, 256K context[87] |
| Inkling | Thinking Machines Lab | Inkling (Jul 15, 2026)[85] | Apache 2.0[86] | MoE, 975B total / 41B active, up to 1M-token context, text, image and audio input; pretrained on 45 trillion tokens[85] |
| Kimi K3 | Moonshot AI | Kimi K3 (Jul 2026)[62], Kimi K3.1 Pro (announced Oct 2026; weights Nov 2, 2026)[65][66] | Kimi K3 License and Kimi K3.1-Pro License (separate agreement for Model-as-a-Service businesses with more than US$20 million in revenue over 12 months)[64][66] | 2.8T total / 104B active MoE with a 1M-token context; Moonshot calls K3 "the first open model to reach 2.8 trillion parameters"[63][62] |
| MiniMax M3 | MiniMax | MiniMax M3 (Jun 2026)[67], MiniMax M3.1 Flash (Oct 19, 2026; weights later)[69] | MiniMax Community License (non-commercial grant; commercial use needs a "Built with" notice and, above US$20 million in yearly revenue, written authorization)[68][70] | M3: about 428B total / 23B active, natively multimodal, 1M-token context with MiniMax Sparse Attention[68][67] |
| Muse Glimmer and Muse Spark Open | Meta Superintelligence Labs | Muse Glimmer (Aug 10, 2026)[82], Muse Spark Open (Oct 15, 2026)[84] | Apache 2.0 (Muse Glimmer)[83]; Muse Model License "with limited commercial terms" (Muse Spark Open)[84] | Glimmer: 30B-class model (29.6B including its vision encoder) sized for one consumer GPU;[82][83] Spark Open: natively multimodal, 816B total / 42B active, 1M-token context[84] |
| Beam | Reflection AI | Beam (announced Oct 5, 2026; weights Oct 16, 2026)[89] | Apache 2.0[89] | Sparse MoE, 501B total / 23B active; text input and output[89][90] |
Open-weight models came close to proprietary frontier models on public leaderboards during 2024 and early 2025. DeepSeek-V3 and DeepSeek-R1 in particular drew attention: DeepSeek described R1's performance as on par with OpenAI's o1.[5] The Stanford AI Index, using the Arena leaderboard, measured the top closed model's lead over the top open-weight model at 8.04% in early January 2024 and 1.70% in February 2025; its 2026 edition put the lead at 3.3% in March 2026, up from 0.5% in August 2024.[28][29]
cost structure
The two delivery models price differently, and the choice has compounding effects on a product's unit economics.
| Cost item | Proprietary API | Self-hosted open weights |
|---|---|---|
| Per-token cost | Direct, posted price (for example, Claude Opus 4.7 at $5 per million input tokens and $25 per million output tokens)[13] | Effectively zero variable cost on top of GPU amortization |
| Infrastructure capex | None (provider absorbs it) | GPU servers, networking, racks, or long reservation contracts on a cloud |
| Inference engineering | None required | Need to operate vLLM, TGI, SGLang, or similar; manage batching, KV cache, autoscaling |
| Idle cost | None | Reserved GPUs cost the same whether or not you serve requests |
| Scaling spikes | Handled by provider, subject to rate limits | You must provision for peak; bursting is harder |
| Compliance and SOC2 | Provided | You inherit the entire stack |
The right answer depends on volume. A thin product that handles a few million tokens per month almost always costs less on a proprietary API than on rented H100s. A heavy product with predictable, sustained throughput (an internal copilot used by 50,000 employees, a search reranker, a customer support classifier) often crosses over and becomes cheaper to self-host. Managed open-weight inference, offered by companies like Together AI, Fireworks AI, and Replicate, sits in between: per-token billing on open models, with the operations handled by the provider. Prices vary widely by model. On November 17, 2026, Together AI listed DeepSeek V4.1 Flash at $0.30 per million input tokens and $1.20 per million output tokens and Kimi K3.1 Pro at $2.00 and $10.00, against Anthropic's list price of $4 and $20 for Claude Opus 5.5.[33][25]
The choice is less binary than it looks: an open-weight model served by a managed inference provider is billed per token like a proprietary API, with the operations handled by the provider.
benchmarks and quality
Four evaluations dominate public discussion of LLM quality:
- LMArena (formerly Chatbot Arena) uses anonymous pairwise human preference voting and reports an Elo rating. It captures perceived helpfulness on open-ended chat better than fixed benchmarks.
- MMLU measures multitask knowledge across 57 subjects, from elementary mathematics to US history, computer science and law.[44] Frontier models now score above 85 percent.
- GPQA is a set of 448 graduate-level multiple-choice science questions written to be "Google-proof": domain experts scored 65% on them, and skilled non-experts with web access 34%.[43] Frontier reasoning models reached human-expert performance during 2024 and 2025.
- SWE-bench evaluates a model's ability to resolve real GitHub issues in open-source repositories. It has become the headline benchmark for coding agents.
The pattern that has emerged is that proprietary models lead the absolute peak by a small margin, while the best open models trail by months rather than years. The frontier reasoning capabilities introduced by OpenAI's o1 in late 2024 were matched by DeepSeek-R1 within a few months, on permissive weights. The gap has not closed steadily: the Stanford AI Index found that it reopened in 2025 after briefly closing in 2024.[29]
pros and cons compared
No summary fully captures every team's situation, but the table below covers the recurring tradeoffs.
| Dimension | Proprietary advantage | Open-weights advantage |
|---|---|---|
| Peak quality | Ahead at the frontier: the top closed model led the top open-weight model by 3.3% on the Arena leaderboard in March 2026[29] | Behind the top closed models by a margin that has both narrowed and widened since 2024[29] |
| Time to first deployment | Provider handles inference; you call an API | Requires GPU provisioning and a serving stack |
| Data handling | Processed on the provider's or a partner cloud's servers under contract terms; OpenAI, for example, says API data is not used for training by default and keeps abuse-monitoring logs for up to 30 days, with zero data retention for eligible customers[39] | Stays on your infrastructure when you self-host; a managed open-weight host is a third party like any other API |
| Fine-tuning | Only what the provider offers, and the tuned weights stay with the provider: as of November 2026 Google offers supervised tuning for several Gemini models, while OpenAI is winding down its fine-tuning platform[37][38] | Any method the license allows: LoRA, QLoRA, full SFT, RLHF, distillation |
| Cost at low volume | Cheaper (no capex, no idle GPUs) | More expensive (you pay for GPUs whether you use them or not) |
| Cost at high volume | More expensive (per-token markup) | Cheaper (amortized GPU cost over many tokens) |
| Vendor lock-in | Higher; switching providers means re-prompting and re-evaluation | Lower; you hold a copy of the weights and can change hosts, within the terms of the model's license |
| Auditability | Limited; the weights cannot be inspected, and an alias can be repointed to a new model, although providers also offer dated snapshots that stay fixed until retirement[46] | High; you can inspect, freeze, and version the model |
| Safety and alignment | Provider-managed; behavior is shaped by their policies | Your responsibility; base models are often less aligned than chat models |
| Long-term availability | Subject to deprecation; older models get retired | Permanent; you keep the weights as long as you keep the storage |
| Operational burden | Low | High; you run the inference stack, monitor it, and patch it |
The "sudden deprecation" risk is worth calling out. OpenAI, Anthropic, and Google all retire older models on published schedules, and applications calibrated to one version often need re-evaluation when forced to upgrade. With open weights, deprecation is your decision.
hybrid strategies
Most mature production stacks use both. A few patterns recur:
- Tiered routing. A cheaper model (often an open-weight one) handles routine requests; a frontier proprietary model handles hard ones. Routers can use simple heuristics, classifiers, or learned policies.
- Proprietary as judge. Many teams use a top-tier closed model to grade outputs from cheaper models during evaluation, even if the cheap model serves production. This decouples evaluation quality from inference cost.
- Distillation. A proprietary teacher generates training data for a smaller open student, which is then deployed. License terms matter here: the Llama 2 and Llama 3 licenses forbid using outputs to improve non-Llama models (Meta dropped the clause from Llama 3.1 onward), and many proprietary providers prohibit using their API outputs to train competing models, which constrains who can do this legally.
- Open base, proprietary fallback. Self-host an open model for the common case and call a proprietary API only when the open model abstains or expresses low confidence.
- Sensitive-data isolation. Use an open model on-premises for any request that touches regulated or confidential data; use proprietary APIs for everything else.
regulatory considerations
Regulation has begun to treat proprietary and open models differently in some respects.
The EU AI Act, whose obligations for general-purpose AI (GPAI) model providers became applicable on August 2, 2025, contains a partial exemption for open-source GPAI models. Providers that release a model under a free and open license, with weights, parameters, architecture, and usage information publicly available, are exempt from the technical documentation and downstream-information requirements that apply to closed providers.[16][17] They still must put in place a copyright-compliance policy and publish a sufficiently detailed summary of the content used for training.[16] The exemption does not apply to models with systemic risk.[16] Under Article 51, a model is presumed to have the high-impact capabilities that lead to that classification when the cumulative compute used to train it exceeds 10^25 floating-point operations.[35] Meta reported that Llama 3.1 405B was pre-trained with 3.8 x 10^25 floating-point operations, above that figure.[36]
US export controls also affect the open-weights ecosystem. Restrictions on advanced GPU exports to China have not stopped Chinese labs from releasing capable open weights (DeepSeek, Qwen, Yi), but they shape what hardware can train them and where the resulting models can be served. Some US legislators have proposed restrictions on the open release of model weights above certain capability thresholds; as of early 2026, no such restriction has become law.
notable historical events
A few episodes shaped the current landscape:
- Llama 1 leak, March 2023. Meta released Llama 1 in February 2023 under a non-commercial research license, restricted to approved researchers. On March 3, 2023, the weights were posted as a torrent on 4chan and spread rapidly.[19][32] Within weeks, Stanford released Alpaca (Llama 1 fine-tuned on instruction-following data) and the Vicuna team produced a chat model that approached GPT-3.5 quality. The leak is widely credited with catalyzing the open-weights ecosystem and pushed Meta to release Llama 2 in July 2023 under a commercial license.[18]
- Mistral 7B, September 2023. Mistral AI, a French startup, released a 7.3B-parameter model under Apache 2.0 and reported that it outperformed Llama 2 13B on every benchmark in its own evaluation.[40]
- DeepSeek-V3 and R1, late 2024 to early 2025. DeepSeek released V3 on December 26, 2024 and R1 on January 20, 2025, and described R1's performance as on par with OpenAI's o1.[5][6] R1 was released under MIT license.[5] The releases triggered a sharp drop in some US AI stocks in late January 2025 and prompted reassessments of the assumed lead held by US labs.
- OpenAI's gpt-oss, August 2025. OpenAI released gpt-oss-120b and gpt-oss-20b under Apache 2.0 on August 5, 2025.[11] These were OpenAI's first open-weight language models since GPT-2 in 2019. The 120B variant runs on a single 80 GB GPU and reaches near-parity with OpenAI's own o4-mini on core reasoning benchmarks, signaling that even the closed labs see strategic value in shipping some open weights.
- Mistral 3 and Large 3, December 2025. Mistral released Mistral Large 3 (41B active / 675B total parameters), which it described as joining "the ranks of frontier instruction-fine-tuned open-source models", and a family of small dense models (3B, 8B, 14B), all under Apache 2.0.[14]
- Gemma 4 moves to Apache 2.0, April 2026. Google released Gemma 4 on April 2, 2026 under "a commercially permissive Apache 2.0 license", writing "You gave us feedback, and we listened."[31] The Gemma Terms of Use, revised on April 1, 2026, now apply only to the models listed in their appendix, from Gemma 1 to Gemma 3n and derivatives such as PaliGemma and ShieldGemma, and point Gemma 4 users to its own license.[95]
- DeepSeek V4, April 2026. DeepSeek released preview weights of DeepSeek-V4-Pro (1.6T total and 49B active parameters) and DeepSeek-V4-Flash (284B and 13B) on April 24, 2026, both with a one-million-token context and under the MIT license.[47][49]
- NVIDIA adopts OpenMDW, May 2026. On May 28, 2026 the Linux Foundation released OpenMDW-1.1 and announced that NVIDIA "will adopt OpenMDW-1.1 across future releases of Cosmos, Isaac GR00T, Ising, and Nemotron open model families".[96] Nemotron 3 Super had shipped in March under the NVIDIA Nemotron Open Model License; Nemotron 3 Ultra (June), Nemotron 3.5 Lightning (August) and Nemotron 3.5 Super VL (October) were released under OpenMDW-1.1.[78][79][80][81]
- Kimi K3, July 2026. Moonshot AI introduced Kimi K3 as "the first open model to reach 2.8 trillion parameters" and released the weights under the Kimi K3 License, which requires Model-as-a-Service businesses with more than US$20 million in revenue over 12 months to sign a separate agreement with Moonshot.[62][64]
- Meta's Muse releases, August and October 2026. Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter model meant to run on a single consumer GPU, under Apache 2.0 on August 10, 2026.[82] On October 15 it published the weights of Muse Spark Open, a natively multimodal model with 816B total and 42B active parameters, under the custom Muse Model License "with limited commercial terms", calling it "the most capable US open-weight model to date".[84]
how to choose
The right answer depends on your constraints, not on philosophy. A few rules of thumb:
- If your data cannot leave your infrastructure for legal, regulatory, or security reasons, self-hosted open weights are the usual answer, though some closed models are also sold for on-premises use: Google made Gemini generally available on its air-gapped Google Distributed Cloud in August 2025.[34]
- If your volume is small and you want to ship quickly, use a proprietary API. The dollar cost is rarely the binding constraint at low volume.
- If your volume is large and your workload is predictable, run an open model on rented or owned GPUs. Per-token economics will dominate.
- If you need maximum capability on the hardest tasks, the top proprietary models still tend to lead, though by less than they used to.
- If you need full control over fine-tuning, alignment, or behavior, open weights give the most control, because a closed provider exposes only the tuning options it chooses to offer.
- For most teams, the best stack uses both: open weights for routine work and the most capable proprietary model for hard requests, evaluation, and synthesis.
The measured gap between the best closed and open-weight models has both narrowed and widened since 2024: the Stanford AI Index put it at 0.5% in August 2024 and 3.3% in March 2026.[29]
Continue reading
- Best Open-Source LLMsAs of November 17, 2026, the highest-scoring open-weight large language model is Moonshot AI's Kimi K3.1 Pro, whose weights Moonshot released on November 4.Listed in See also
- LLaMALLaMA is the original name of a family of large language models developed by Meta AI.Mentioned 46 times above
- DeepSeekDeepSeek is a Chinese artificial intelligence company based in Hangzhou. It was founded in 2023 by Liang Wenfeng, who had previously co-founded the quantitative investment firm High-Flyer.Mentioned 27 times above
- Mistral AIMistral AI is a French artificial intelligence company that develops large language models, multimodal models, software for building and operating AI systems, and computing infrastructure.Mentioned 24 times above
- Apache License 2.0The Apache License, Version 2.0 is a permissive open source software license published by the Apache Software Foundation (ASF) in January 2004 and approved by the ASF board on January 21, 2004.Mentioned 23 times above
- OpenAIOpenAI is an American artificial intelligence research and deployment organization.Mentioned 14 times above
More to read
see also
references
- ^1 ^2Open Source Initiative. "The Open Source AI Definition 1.0." October 2024. opensource.org/...open-source-ai-definition
- ^1 ^2Meta. "Llama 3 Community License Agreement." llama.com/...license
- ^1 ^2 ^3Meta. "Llama 3.1 Community License Agreement." llama.com/...license
- ^Meta. "The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation." April 5, 2025. ai.meta.com/...llama-4-multimodal-intelligence
- ^1 ^2 ^3 ^4DeepSeek. "DeepSeek-R1 Release." January 20, 2025. api-docs.deepseek.com/...news250120
- ^1 ^2DeepSeek-AI. "DeepSeek-V3 Technical Report." December 2024. arxiv.org/...2412.19437
- ^1 ^2Qwen Team. "Qwen3: Think Deeper, Act Faster." April 29, 2025. qwenlm.github.io/...qwen3
- ^1 ^2Google DeepMind. "Gemma 3 Technical Report." March 2025. storage.googleapis.com/...Gemma3Report.pdf
- ^Google. "Introducing Gemini 3." November 18, 2025. blog.google/...gemini-3
- ^OpenAI. "Introducing GPT-5." August 7, 2025. openai.com/...introducing-gpt-5
- ^1 ^2OpenAI. "Introducing gpt-oss." August 5, 2025. openai.com/...introducing-gpt-oss
- ^OpenAI. "Hello GPT-4o." May 13, 2024. openai.com/...hello-gpt-4o
- ^1 ^2Anthropic. "Introducing Claude Opus 4.7." April 16, 2026. anthropic.com/...claude-opus-4-7
- ^1 ^2Mistral AI. "Introducing Mistral 3." December 2, 2025. mistral.ai/...mistral-3
- ^1 ^2Allen Institute for AI. "OLMo 2: The best fully open language model to date." allenai.org/...olmo2
- ^1 ^2 ^3European Union. "Article 53: Obligations for Providers of General-Purpose AI Models." EU Artificial Intelligence Act. artificialintelligenceact.eu/...53
- ^European Commission. "Guidelines for providers of general-purpose AI models." digital-strategy.ec.europa.eu/...es-gpai-providers
- ^Wikipedia. "Llama (language model)." en.wikipedia.org/...Llama_(language_model)
- ^The Register. "LLaMA drama as Meta's mega language model files leak." March 8, 2023. theregister.com/...meta_llama_ai_leak
- ^LMArena. "Arena Leaderboard." huggingface.co/...arena-leaderboard
- ^Hugging Face. "Welcome Llama 4 Maverick & Scout on Hugging Face." April 2025. huggingface.co/...llama4-release
- ^Simon Willison. "Meta AI release Llama 3.3." December 6, 2024. simonwillison.net/...llama-33
- ^OpenAI. "GPT-6 Astra: A new generation of intelligence." September 3, 2026. openai.com/...gpt-6-astra
- ^Anthropic. "Introducing Claude Fable 5.2 and Claude Mythos 5.2." November 9, 2026. anthropic.com/claude-fable-and-mythos-5-2
- ^1 ^2Anthropic. "Introducing Claude Opus 5.5." September 22, 2026. anthropic.com/claude-opus-5-5
- ^Koray Kavukcuoglu. "Gemini 4 Carbon: our next frontier model." Google, October 28, 2026. blog.google/...introducing-gemini-4-carbon
- ^SpaceXSI (formerly xAI). "Introducing Grok 4.7." September 21, 2026. x.ai/...grok-4-7
- ^Stanford HAI. "Technical Performance." The 2025 AI Index Report. hai.stanford.edu/...technical-performance
- ^1 ^2 ^3 ^4 ^5Stanford HAI. "Technical Performance." The 2026 AI Index Report. hai.stanford.edu/...technical-performance
- ^1 ^2Meta. "Introducing Llama 3.1: Our most capable models to date." July 23, 2024. ai.meta.com/...meta-llama-3-1
- ^1 ^2 ^3 ^4 ^5 ^6Google. "Gemma 4: Byte for byte, the most capable open models." April 2, 2026. blog.google/...gemma-4
- ^James Vincent. "Meta's powerful AI language model has leaked online - what happens now?" The Verge, March 8, 2023. theverge.com/...age-model-llama-leak-online-misuse
- ^Together AI. "Pricing." Accessed November 17, 2026. together.ai/pricing
- ^Google Cloud. "Run Gemini anywhere, including on-premises, with Google Distributed Cloud." August 28, 2025. cloud.google.com/...gemini-is-now-available-anywhere
- ^European Union. "Article 51: Classification of General-Purpose AI Models as General-Purpose AI Models with Systemic Risk." EU Artificial Intelligence Act. artificialintelligenceact.eu/...51
- ^Meta. "The Llama 3 Herd of Models." arXiv:2407.21783, July 2024. arxiv.org/...2407.21783
- ^OpenAI. "Model optimization." OpenAI API documentation. Accessed November 17, 2026. developers.openai.com/...model-optimization
- ^Google Cloud. "Tune Gemini models with supervised fine-tuning." Accessed November 17, 2026. docs.cloud.google.com/...use
- ^OpenAI. "Data controls in the OpenAI platform." OpenAI API documentation. Accessed November 17, 2026. developers.openai.com/...your-data
- ^Mistral AI. "Mistral 7B." September 27, 2023. mistral.ai/...announcing-mistral-7b
- ^Kyle Wiggers. "We finally have an 'official' definition for open source AI." TechCrunch, October 28, 2024. techcrunch.com/...al-definition-for-open-source-ai
- ^MUFG Capital Markets Strategy. "The AI Weekly: Global AI LM Rankings in April 2026" (Arena data as of April 8, 2026). April 10, 2026. mufgamericas.com/..._Global_AI_Language_Models.pdf
- ^David Rein et al. "GPQA: A Graduate-Level Google-Proof Q&A Benchmark." arXiv:2311.12022, November 2023. arxiv.org/...2311.12022
- ^Dan Hendrycks et al. "Measuring Massive Multitask Language Understanding." arXiv:2009.03300, September 2020. arxiv.org/...2009.03300
- ^DeepSeek. "DeepSeek License Agreement, Version 1.0" (LICENSE-MODEL in the DeepSeek-V3 repository). huggingface.co/...LICENSE-MODEL
- ^OpenAI. "GPT-4.1 Model." OpenAI API documentation. Accessed November 17, 2026. developers.openai.com/...gpt-4.1
- ^1 ^2 ^3DeepSeek. "DeepSeek V4 Preview Release." DeepSeek API Docs, April 24, 2026. api-docs.deepseek.com/...news260424
- ^DeepSeek. "DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient." DeepSeek API Docs, September 10, 2026. api-docs.deepseek.com/...news260910
- ^1 ^2 ^3DeepSeek-AI. "DeepSeek-V4-Pro" model card (license: MIT). Hugging Face. Accessed November 17, 2026. huggingface.co/...DeepSeek-V4-Pro
- ^1 ^2DeepSeek-AI. "DeepSeek-V4.1-Flash" model card (license: MIT). Hugging Face. Accessed November 17, 2026. huggingface.co/...DeepSeek-V4.1-Flash
- ^1 ^2Qwen Team. "QwenLM/Qwen3.8" GitHub repository (README news log). Accessed November 17, 2026. github.com/...Qwen3.8
- ^1 ^2Qwen. "Qwen3.5-397B-A17B" model card (license: Apache 2.0). Hugging Face. Accessed November 17, 2026. huggingface.co/...Qwen3.5-397B-A17B
- ^Qwen. "Qwen3.6-35B-A3B" model card (license: Apache 2.0). Hugging Face. Accessed November 17, 2026. huggingface.co/...Qwen3.6-35B-A3B
- ^Qwen. "Qwen3.8-27B" model card (license: Apache 2.0). Hugging Face. Accessed November 17, 2026. huggingface.co/...Qwen3.8-27B
- ^Qwen. "Qwen3.8-2.4T-A95B" model card (license: qwen3.8-max). Hugging Face. Accessed November 17, 2026. huggingface.co/...Qwen3.8-2.4T-A95B
- ^1 ^2 ^3 ^4Qwen. "Qwen3.8-Max License." Hugging Face, Qwen/Qwen3.8-2.4T-A95B. Accessed November 17, 2026. huggingface.co/...LICENSE
- ^Z.ai. "New Released" (model release notes). Z.ai Developer Documentation. Accessed November 17, 2026. docs.z.ai/...new-released
- ^1 ^2Z.ai. "GLM-5" model card (license: MIT). Hugging Face. Accessed November 17, 2026. huggingface.co/...GLM-5
- ^Z.ai. "GLM-5.1" model card (license: MIT). Hugging Face. Accessed November 17, 2026. huggingface.co/...GLM-5.1
- ^Z.ai. "GLM-5.2" model card (license: MIT). Hugging Face. Accessed November 17, 2026. huggingface.co/...GLM-5.2
- ^1 ^2 ^3 ^4Z.ai. "GLM-5.3 License." Hugging Face, zai-org/GLM-5.3. Accessed November 17, 2026. huggingface.co/...LICENSE
- ^1 ^2 ^3Moonshot AI. "Kimi K3 Tech Blog: Open Frontier Intelligence." Kimi, July 2026. kimi.com/...kimi-k3
- ^Moonshot AI. "Kimi-K3" model card (license: Kimi K3 License). Hugging Face. Accessed November 17, 2026. huggingface.co/...Kimi-K3
- ^1 ^2 ^3Moonshot AI. "Kimi K3 License." Hugging Face, moonshotai/Kimi-K3. Accessed November 17, 2026. huggingface.co/...LICENSE
- ^Moonshot AI. "Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence." Kimi, October 2026. kimi.ai/...kimi-k3-1-pro
- ^1 ^2 ^3 ^4 ^5Moonshot AI. "Kimi K3.1-Pro License." Hugging Face, moonshotai/Kimi-K3.1-Pro. Accessed November 17, 2026. huggingface.co/...LICENSE
- ^1 ^2MiniMax. "MiniMax M3: Frontier Coding, 1M Context, Native Multimodality, All in One Model." MiniMax Research, June 2026. minimax.io/...minimax-m3
- ^1 ^2MiniMax. "MiniMax-M3" model card (license: MiniMax Community License). Hugging Face. Accessed November 17, 2026. huggingface.co/...MiniMax-M3
- ^MiniMax. "MiniMax M3.1 Flash: Redefining Agent Performance and Cost." MiniMax Research, October 19, 2026. minimax.io/...minimax-m31f
- ^MiniMax. "MiniMax Community License" (MiniMax-M3.1-Flash). Hugging Face. Accessed November 17, 2026. huggingface.co/...LICENSE
- ^1 ^2 ^3Mistral AI. "Introducing Mistral Small 4." March 16, 2026. mistral.ai/...mistral-small-4
- ^1 ^2Mistral AI. "Remote agents in Vibe. Powered by Mistral Medium 3.5." Mistral AI news. Accessed November 17, 2026. mistral.ai/...vibe-remote-agents-mistral-medium-3-5
- ^Mistral AI. "Modified MIT License" (Mistral-Medium-3.5-128B). Hugging Face. Accessed November 17, 2026. huggingface.co/...LICENSE
- ^Mistral AI. "Introducing Mistral Large 4." October 6, 2026. mistral.ai/...mistral-large-4
- ^Mistral AI. "Mistral Large 4 (Le Chonk) is now open weight." October 29, 2026. mistral.ai/...mistral-large-4-oss
- ^Mistral AI. "Mistral-Large-4-1T-A52B" model card. Hugging Face. Accessed November 17, 2026. huggingface.co/...Mistral-Large-4-1T-A52B
- ^1 ^2 ^3 ^4Mistral AI. "Mistral Ecosystem Open-Weight License." Hugging Face, mistralai/Mistral-Large-4-1T-A52B. Accessed November 17, 2026. huggingface.co/...LICENSE
- ^1 ^2 ^3 ^4NVIDIA. "NVIDIA-Nemotron-3-Super-120B-A12B-BF16" model card. Hugging Face. Accessed November 17, 2026. huggingface.co/...-Nemotron-3-Super-120B-A12B-BF16
- ^1 ^2 ^3 ^4 ^5NVIDIA. "NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16" model card. Hugging Face. Accessed November 17, 2026. huggingface.co/...-Nemotron-3-Ultra-550B-A55B-BF16
- ^1 ^2 ^3 ^4 ^5NVIDIA. "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16" model card. Hugging Face. Accessed November 17, 2026. huggingface.co/...otron-3.5-Lightning-30B-A3B-BF16
- ^1 ^2 ^3 ^4NVIDIA. "NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16" model card. Hugging Face. Accessed November 17, 2026. huggingface.co/...tron-3.5-Super-VL-120B-A12B-BF16
- ^1 ^2 ^3 ^4Meta Superintelligence Labs. "Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device." Meta AI Research, August 10, 2026. research.meta.ai/...use-glimmer-open-agentic-model
- ^1 ^2Meta. "Muse-Glimmer-30B" model card (license: Apache 2.0). Hugging Face. Accessed November 17, 2026. huggingface.co/...Muse-Glimmer-30B
- ^1 ^2 ^3 ^4Meta Superintelligence Labs. "Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date." Meta AI Research, October 15, 2026. research.meta.ai/...muse-spark-open
- ^1 ^2Thinking Machines Lab. "Inkling: Our Open-Weights Model." July 15, 2026. thinkingmachines.ai/...introducing-inkling
- ^Thinking Machines Lab. "Inkling" model card (license: Apache 2.0). Hugging Face. Accessed November 17, 2026. huggingface.co/...Inkling
- ^1 ^2Tencent. "腾讯混元 Hy3,正式发布" (Tencent Hunyuan Hy3 officially released). Tencent News, July 6, 2026. news.qq.com/...20260706A07GV200
- ^Tencent. "Hy3" model card (license: Apache 2.0). Hugging Face. Accessed November 17, 2026. huggingface.co/...Hy3
- ^1 ^2 ^3Reflection AI. "Beam is now available." October 16, 2026. reflection.ai/...beam-is-available
- ^Reflection AI. "Beam-1.0-501B-A23B" model card (license: Apache 2.0). Hugging Face. Accessed November 17, 2026. huggingface.co/...Beam-1.0-501B-A23B
- ^NVIDIA. "Nemotron-4-340B-Instruct" model card (license: NVIDIA Open Model License). Hugging Face. Accessed November 17, 2026. huggingface.co/...Nemotron-4-340B-Instruct
- ^NVIDIA. "NVIDIA Open Model License Agreement." Last modified October 24, 2025. nvidia.com/...nvidia-open-model-license
- ^Cohere Labs. "c4ai-command-a-03-2025" model repository (license: cc-by-nc-4.0). Hugging Face. Accessed November 17, 2026. huggingface.co/...c4ai-command-a-03-2025
- ^1 ^2G7 Digital and Technology Ministers. "G7 Vision on AI Openness Opportunities and Shared Language." Paris, May 29, 2026. G7 Information Centre, University of Toronto. g7.utoronto.ca/...260529-vision
- ^Google. "Gemma Terms of Use." Last modified April 1, 2026. ai.google.dev/...terms
- ^The Linux Foundation. "Linux Foundation Releases OpenMDW-1.1; NVIDIA Adopts OpenMDW for Cosmos, Isaac GR00T, Ising and Nemotron, AI Model Families." May 28, 2026. linuxfoundation.org/...-nemotron-ai-model-families
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
10 revisions · v10 · 6,458 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: November 2026 update (11 open-weight rows, license taxonomy, 2026 events) checked by an independent verifier: 96 claims against 62 sources, 6 minor fixes applied.
Cite this page: AI Wiki. "Proprietary vs. Open Source Large Language Models (LLMs)." aiwiki.ai, updated 17 Nov 2026, fact-checked 17 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/proprietary_vs_open_source_large_language_models_llms



