Open-Weight LLM License Comparison

As of July 2026, the open-weight LLMs that are genuinely free for commercial use with no strings attached are the ones released under standard, OSI-approved licenses: Apache-2.0 (Alibaba's Qwen3, most of Mistral's open models, OpenAI's gpt-oss, and Ai2's OLMo 2) and MIT (DeepSeek's V3 and R1, Zhipu/Z.ai's GLM-4.6, and Microsoft's Phi-4). With those licenses you can download the weights, self-host, fine-tune, ship a commercial product, and keep your outputs, with essentially no user or revenue cap ^[5][6].

Is Llama commercially free? Mostly yes, but Llama is not open source. The Llama 4 Community License lets you use it commercially only if the products or services you offer had fewer than 700 million monthly active users (MAU) in the calendar month before Llama 4's release; above that you must request a separate license from Meta, granted at its sole discretion ^[1]. It also forces you to display "Built with Llama" and to put "Llama" at the start of any derivative model's name ^[1]. Google's Gemma 3 is similar: free to use and redistribute, but bound by a Prohibited Use Policy that Google can enforce remotely ^[3][4]. The one-line rule: Apache-2.0 and MIT models are open source; Llama, Gemma, Mistral's research licenses, and OpenRAIL are "open weight but restricted."

Which open models are truly free for commercial use?

Ranked by how few strings are attached:

1. Tier 1, no strings (OSI open source): anything under Apache-2.0 or MIT. Keep the license notice and you are done. This covers Qwen3, DeepSeek-R1/V3, GLM-4.6, gpt-oss, most Mistral open models, OLMo 2, and Phi-4.
2. Tier 2, free for almost everyone but with conditions: Llama 4 (fine below 700M MAU, plus branding and naming rules) and Gemma 3 (fine, but you inherit Google's use restrictions).
3. Tier 3, conditional or non-commercial: the Mistral Research License (research only), Mistral's "modified MIT" on Devstral 2 (free unless your company clears about USD 20M/month in revenue), and OpenRAIL (free, but with behavioral use bans that follow every derivative).

The quick per-model answer:

Scroll sideways for more →

Open-weight license comparison

Each license below governs the weights and, usually, the accompanying code. Model outputs are treated separately (see "Do you own the model's outputs?").

Scroll sideways for more →

Last verified: July 2026. License names use the canonical SPDX identifier where one exists.

Is Llama commercially free? The 700 million MAU rule

Yes for the overwhelming majority of users, and no for a handful of hyperscalers. The Llama 4 Community License grants a "non-exclusive, worldwide, non-transferable and royalty-free limited license" to "use, reproduce, distribute, copy, create derivative works of, and make modifications to the Llama Materials" ^[1]. The catch is the MAU clause: "If, on the Llama 4 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee's affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta" ^[1]. In practice only companies at the scale of the largest social and cloud platforms cross that line; a startup, an enterprise, or a solo developer can use Llama 4 in production for free.

Two obligations remain even below the cap. First, redistribution requires shipping a copy of the agreement and prominently displaying "Built with Llama" on a related website, interface, or documentation. Second, if you use Llama or its outputs to train or fine-tune another distributed model, you "shall also include 'Llama' at the beginning of any such AI model name" ^[1]. Use must also follow the Llama 4 Acceptable Use Policy ^[2]. Because these terms restrict who may use the model and how derivatives are named, the license is not open source under the Open Source Initiative definition ^[15].

What do the Gemma Terms of Use restrict?

Gemma is free to use, modify, and redistribute for commercial purposes, and there is no MAU or revenue cap ^[3]. The constraints are behavioral and administrative. If you distribute Gemma or a model derivative, you must pass along Google's use restrictions as an enforceable term, include the Prohibited Use Policy, provide the full agreement to downstream recipients, and mark any files you changed ^[3]. Google also reserves the right to "restrict (remotely or otherwise) usage of any of the Gemma Services" it reasonably believes violate the agreement ^[3], and it can update the Prohibited Use Policy over time ^[4]. Creating derivatives, including through distillation, is explicitly allowed ^[3]. Like the Llama license, Gemma is open weight but not OSI open source, because it carries field-of-use restrictions.

Which Mistral models are Apache-2.0 and which are restricted?

Most current Mistral open models are Apache-2.0 ^[7]. The shift back to permissive licensing began with Mistral Small 3 on 30 January 2025, and subsequent open releases, including Mistral Small 3.1/3.2, the Ministral 3 family, Magistral Small, Devstral Small, and Mistral Large 3, have shipped under Apache-2.0 ^[7]. Two categories are restricted. The older Mistral Research License (MRL) applies to models such as Mistral Large 2 (2407) and Pixtral Large: free for research and non-commercial use, with a paid license required for production ^[7][8]. Newer still, a "modified MIT" license on the flagship Devstral 2 (123B) keeps MIT-style freedom but adds a revenue gate: a company with global consolidated monthly revenue above roughly USD 20M must obtain a commercial license or route through Mistral Studio ^[7]. Note the pair split: Devstral 2 (123B) is modified MIT, while Devstral Small 2 (24B) is plain Apache-2.0 ^[7].

Qwen and DeepSeek show the same "flagship open, top model held back" pattern. Qwen3 open weights are Apache-2.0, but earlier Qwen models used a source-available Qwen License or a non-commercial Qwen Research License, and Qwen-Max is proprietary and API-only. DeepSeek tightened, then loosened: DeepSeek-V3 first shipped under a custom model license modeled partly on OpenRAIL, then was relicensed to plain MIT in March 2025, and DeepSeek-R1 has been MIT since its January 2025 launch ^[9][10].

Open source vs open weight: what is the real difference?

"Open weight" means you can download and run the model. "Open source," in the Open Source Initiative sense, additionally means the license may "not restrict anyone from making use of the program in a specific field of endeavor" and imposes no discrimination against users or uses ^[15]. That distinction is exactly where community licenses fall short: Llama's MAU gate and naming rule, Gemma's Prohibited Use Policy, Mistral's revenue gate, and OpenRAIL's behavioral bans are all field-of-use or user restrictions, so none of them are OSI-approved, even though the weights are downloadable. OpenRAIL (the Responsible AI License family behind BLOOM, Stable Diffusion, and BigCode's StarCoder) is the clearest case: it permits free use, redistribution, and commercialization, but embeds behavioral use restrictions (no surveillance, disinformation, discrimination, or weapons uses) that must propagate unchanged or stricter to every derivative ^[14]. The OSI, and most corporate legal teams, classify it as source-available rather than open source ^[15]. Only Apache-2.0 and MIT among the licenses here clear the bar. For fully open source down to the training data and code, OLMo 2 from Ai2 is the reference point ^[16].

Do you own the model's outputs?

Across every license here, the provider does not claim ownership of the text, code, or images you generate: outputs are yours to use and, in most jurisdictions, are not themselves copyrightable. The nuances are contractual, not proprietary. Under Apache-2.0 and MIT there are no strings on outputs at all. Under the Llama 4 license, outputs are yours, but if you use them to train a distributed model, that model must be named "Llama-..." ^[1]. Under Gemma, Google claims no rights to outputs, yet their generation and use must comply with the Prohibited Use Policy ^[3][4]. Under OpenRAIL, the behavioral restrictions extend to outputs as well ^[14]. Under the Mistral Research License, outputs from non-commercial use may not be repurposed commercially without a paid license ^[8].

Is Llama 4 Meta's last open release?

As of mid-2026, yes in practice. Llama 4 Scout and Maverick, released in April 2025, remain Meta's most recent open-weight models; Meta has not shipped new Llama weights since ^[11][17]. In December 2025 Bloomberg reported that Meta planned to move to closed-source distribution for future frontier LLMs, and the first model from Meta Superintelligence Labs, Muse Spark (April 2026), launched as a proprietary, API-only model ^[17][18]. Meta has said it "hopes to open-source future versions," and reporting describes reduced open derivatives of its "Avocado" and "Mango" projects arriving eventually, but those would omit capabilities and are not committed to a date ^[17]. The takeaway for licensing decisions: treat Llama 4 as the endpoint of Meta's open-weight line for now, and expect the most permissive frontier weights to come from Alibaba (Qwen), DeepSeek, Zhipu/Z.ai, and Mistral rather than Meta.

References

1. Meta, "Llama 4 Community License Agreement," github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE (also llama.com/llama4/license). Accessed July 2026. ↩
2. Meta, "Llama 4 Acceptable Use Policy," llama.com/llama4/use-policy. Accessed July 2026. ↩
3. Google, "Gemma Terms of Use," ai.google.dev/gemma/terms. Accessed July 2026. ↩
4. Google, "Gemma Prohibited Use Policy," ai.google.dev/gemma/prohibited_use_policy. Accessed July 2026. ↩
5. Apache Software Foundation, "Apache License, Version 2.0," apache.org/licenses/LICENSE-2.0. ↩
6. Open Source Initiative, "The MIT License," opensource.org/license/mit. ↩
7. Mistral AI Help Center, "Under which license are Mistral's open models available?" help.mistral.ai. Accessed July 2026. ↩
8. Mistral AI, "Mistral Research License" and "Mistral AI Non-Production License (MNPL)," mistral.ai/news/mistral-ai-non-production-license-mnpl. ↩
9. DeepSeek, "DeepSeek-R1 LICENSE (MIT)," github.com/deepseek-ai/DeepSeek-R1/blob/main/LICENSE; "DeepSeek-R1 Release," api-docs.deepseek.com/news/news250120. ↩
10. SiliconANGLE, "DeepSeek releases improved DeepSeek-V3 model under MIT license," 24 March 2025. ↩
11. Alibaba Qwen, "Qwen3 LICENSE (Apache-2.0)," github.com/QwenLM/Qwen3. ↩
12. Z.ai / Zhipu AI, "GLM-4.6 model card (MIT license)," Hugging Face zai-org/GLM-4.6.
13. OpenAI, "Introducing gpt-oss" (Apache-2.0), openai.com/index/introducing-gpt-oss, 5 August 2025.
14. Hugging Face and licenses.ai, "OpenRAIL: Towards open and responsible AI licensing frameworks"; "BigScience OpenRAIL-M License." ↩
15. Open Source Initiative, "The Open Source Definition" and "Open Source AI Definition"; OSI analysis of OpenRAIL and model-weight licenses. ↩
16. Allen Institute for AI (Ai2), "OLMo 2" (Apache-2.0, open data and training code), allenai.org/olmo. ↩
17. SiliconANGLE, "Report: Meta developing open-source versions of upcoming AI models," 6 April 2026; Bloomberg reporting on Meta's closed-source shift, December 2025. ↩
18. Meta, "Introducing Muse Spark: Meta Superintelligence Labs," about.fb.com/news, April 2026. ↩