The state of open-weight licensing, mapped

TL;DR

• Most flagship open-weight releases since mid-2025 ship under Apache 2.0 or MIT: OpenAI’s gpt-oss, Qwen3 and Qwen3.5, Mistral 3, Gemma 4, DeepSeek R1 and V4, Phi-4 and Meta’s Muse Glimmer, according to their model cards and licence files.
• The holdouts carry scale clauses. Llama 3.x and Llama 4 require a separate licence above 700 million monthly active users. Moonshot’s Kimi K2 requires “Kimi K2” on the interface above 100 million users or 20 million dollars of monthly revenue, and Kimi K3 adds a revenue gate for hosting businesses.
• A permissive weight licence does not make a model Open Source AI under the OSI’s definition, which also requires code and “sufficiently detailed information about the data used to train the system”.
• The EU AI Act’s Article 53(2) exemption removes two of four provider duties for free and open-source models, and only for models without systemic risk.

Every open-weight model comes with a document that decides whether you can ship a product on it, and it is rarely the one people read. This survey reads it. We fetched the licence text and model card for each notable open-weight family as of September 2026, sorted them into five families, and pulled out the one clause in each that most often changes a build decision. The finding is that the field has moved: permissive licences, a minority in 2024, now cover most of the largest new releases, while restrictive terms have concentrated in a few vendors and in older models that are still widely deployed.

Five licence families

The table sorts licences by what they let a downstream company do, not by what the vendor calls them. “OSI approved” means the licence text is on the Open Source Initiative’s list; whether the whole system meets the OSI’s separate Open Source AI Definition is covered below.

Permissive licences as the new default

The clearest change since 2024 is who now uses Apache 2.0 and MIT. OpenAI’s gpt-oss model card, dated 5 August 2025, says the 116.8 billion and 20.9 billion parameter models are “available under the Apache 2.0 license and our gpt-oss usage policy”. That usage policy is two sentences long; its only obligation is that “you agree to comply with all applicable law”.

Alibaba’s Qwen3 launch post of 29 April 2025 says the models are released “under Apache 2.0 license”, and the model cards for Qwen3-235B-A22B and the February 2026 Qwen3.5-397B-A17B carry the apache-2.0 tag. The older Qwen LICENSE AGREEMENT, dated 19 September 2024, still governs Qwen2.5-72B: “If you are commercially using the Materials, and your product or service has more than 100 million monthly active users, you shall request a license from us.” Check which generation you are downloading.

Mistral’s Mistral 3 announcement of 2 December 2025 covers Mistral Large 3 (675 billion total, 41 billion active parameters) and Ministral 3 at 14B, 8B and 3B, and says “All models are released under the Apache 2.0 license.” Google’s Gemma 4 post of 2 April 2026 says “Gemma 4 is released under a commercially permissive Apache 2.0 license”, and the Gemma 4 31B card agrees, though Google’s licence page still links a separate prohibited-use document.

DeepSeek’s R1 model card states that “the model weights are licensed under the MIT License” and that the series “allow for any modifications and derivative works, including, but not limited to, distillation for training other LLMs.” The 2026 DeepSeek-V4-Pro card, 1.6 trillion parameters with 49 billion active, is also MIT. Microsoft’s Phi-4, released 12 December 2024, is MIT. xAI’s Grok-1 repository says its weights “are licensed under the Apache 2.0 license.” And Meta, whose Llama terms defined the community-licence family, described its 30 billion parameter Muse Glimmer on 10 August 2026 as “open sourcing the model weights under a permissive Apache 2.0 license”.

NVIDIA’s Nemotron 3 Ultra card points to the OpenMDW License 1.1, updated 27 May 2026, which grants permission “to deal in the Model Materials without restriction” and “does not impose any restrictions or obligations with respect to any use, modification, or sharing of any outputs”. We could not confirm whether OpenMDW is OSI approved.

Scale clauses and naming rules

The Llama Community License is the reference case. The Llama 4 text, effective 5 April 2025, says: “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”. The Llama 3.3 text of 6 December 2024 is identical on this point.

The naming rule is the clause most teams miss. Both versions require you to “prominently display ‘Built with Llama’ on a related website, user interface, blogpost, about page, or product documentation”, and continue: “If you use the Llama Materials or any outputs or results of the Llama Materials to create, train, fine tune, or otherwise improve an AI model, which is distributed or made available, you shall also include ‘Llama’ at the beginning of any such AI model name”. Note “outputs or results”: a model distilled from Llama outputs falls under the naming rule. The licence incorporates the Llama 4 Acceptable Use Policy, which states that for multimodal models “the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union”, though not to “end users of a product or service that incorporates any such multimodal models.”

Moonshot’s Kimi K2 licence is MIT plus one paragraph: “if the Software (or any derivative works thereof) is used for any of your commercial products or services that have more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue, you shall prominently display ‘Kimi K2’ on the user interface”. The January 2026 Kimi K2.5 uses the same modified MIT. The Kimi K3 License, which shipped with the 2.8 trillion parameter K3 in mid-2026, keeps that display clause and adds a gate: “If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose.” That clause targets inference hosts, and it moves Moonshot from attribution to a commercial condition.

Prohibited-use policies and remote control

The third family grants commercial use but attaches a policy the vendor controls. Google’s Gemma Terms of Use, last modified 1 April 2026 and still governing Gemma 1 through 3, say “You must not use any of the Gemma Services: for the restricted uses set forth in the Gemma Prohibited Use Policy”, require distributors to pass those restrictions on “as an enforceable provision”, and add that “Google reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement.” The Prohibited Use Policy covers, among other things, content that “impersonates an individual”.

xAI published Grok 2 weights in August 2025. The LICENSE file in that repository is titled xAI Community License Agreement, last updated 4 November 2025. It permits commercial use “solely if you and your affiliates abide by all of the guardrails provided in xAI’s Acceptable Use Policy”, requires “Powered by xAI” to be displayed, and states: “You may not use the Materials, derivatives, or outputs (including generated data) to train, create, or improve any foundational, large language, or general-purpose AI models”. As of our check, xAI’s Hugging Face organisation lists only Grok-1 and Grok-2.

NVIDIA’s Open Model License, last modified 24 October 2025, governs Nemotron Nano 9B v2 and the December 2025 Nemotron 3 Nano. It is commercially permissive but says that if you “bypass, disable, reduce the efficacy of, or circumvent any technical limitation, safety guardrail or associated safety guardrail hyperparameter” without “a substantially similar Guardrail appropriate for your use case, your rights under this Agreement will automatically terminate.” The NVIDIA Open Model Agreement of 2 April 2026, used by Nemotron 3 Nano Omni, drops that clause and says simply “Works are commercially usable.”

DeepSeek belongs here only for its older weights. The original DeepSeek-V3 of December 2024 used the DeepSeek License Agreement, version 1.0 of 23 October 2023, which lists use-based restrictions such as “For military use in any way” and reserves “the right to restrict (remotely or otherwise) usage of the Model in violation of this License.” DeepSeek-V3-0324 and everything after it moved to MIT.

Research-only and responsible AI terms

Mistral keeps two non-commercial licences for models it has not relicensed. The Mistral AI Research License limits use to “personal, scientific or academic research, and (b) for non-profit and non-commercial purposes”, and applies to Mistral Large 2411 and Ministral 8B 2410. The Mistral AI Non-Production License, which covers Codestral 22B, permits use only “for testing, research, Personal, or evaluation purposes in Non-Production Environments”. Both say that otherwise “You must request a license from Mistral AI.”

The OpenRAIL family predates all of this. Hugging Face’s 2022 explainer describes licences that combine “royalty free access and flexible downstream use” with “use-based restrictions for the use of the licensed AI artifact in identified critical scenarios”, which derivatives must carry forward. BLOOM, released 11 July 2022, ships under BigScience BLOOM RAIL 1.0. No 2025 or 2026 flagship in our survey uses a RAIL licence, though its structure lives on in the Gemma and DeepSeek use-restriction lists.

Model by model

Dates are the licence’s own effective date where it gives one, otherwise the release date on the model card. Vendors update licence pages in place, so treat every entry as a snapshot.

Open source after OSAID 1.0

The Open Source Initiative released the Open Source AI Definition 1.0 on 28 October 2024. Its first freedom is to “Use the system for any purpose and without having to ask for permission”. A scale clause or a prohibited-use list fails that test on its face, so no licence outside the permissive family qualifies. The OSI said so about Llama in its February 2025 post: the licence “Fails at freedom 0, the freedom to use the model for any purpose” and “Fails at the Open Source Definition point 6, restricts fields of endeavour”.

The harder test is what the definition asks for beyond the licence. It requires “The complete source code used to train and run the system” and Data Information, defined as “Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system.” That is where Apache 2.0 flagships fall short. A permissive licence makes the weights open source software in the ordinary sense; it does not make the system Open Source AI unless the training data is documented to that standard, and no frontier-scale release in our survey publishes data at that level. The nearest approach is NVIDIA, whose Nemotron 3 announcement lists released pretraining and post-training datasets, though we did not assess them against the OSAID bar. Our guide to reading a model card like an auditor covers what to look for in the data section.

The EU AI Act exemption and its edges

The AI Act’s obligations for general-purpose model providers applied from 2 August 2025. Article 53(1) requires technical documentation for regulators, documentation for downstream providers, a copyright policy and “a sufficiently detailed summary about the content used for training”. Article 53(2) then says the first two “shall not apply to providers of AI models that are released under a free and open-source licence that allows for the access, usage, modification, and distribution of the model, and whose parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available. This exception shall not apply to general-purpose AI models with systemic risks.”

Three limits follow. First, the copyright policy and training-data summary duties apply to every provider, open or not. Second, Recital 103 says components “provided against a price or otherwise monetised, including through the provision of technical support or other services” do not benefit, while “making AI components available through open repositories should not, in itself, constitute a monetisation”. Third, the Commission’s guidelines of 18 July 2025 interpret the licence condition. According to Hugging Face’s summary of those guidelines, research-only and no-commercial-use licences do not qualify, while “specific, proportionate, and safety-oriented usage restrictions may be permissible” in high-risk domains. On that reading, the Mistral research licences and the Llama user cap fall outside the exemption, and the largest permissive models may still be caught by the systemic-risk carve-out.

Licence first, benchmark second

Two notes close the survey. First, licences on the same weights change: DeepSeek moved V3 to MIT within months, NVIDIA has used three documents in a year, and Google rewrote its Gemma terms and then set them aside for Gemma 4. Pin the licence version you accepted and re-check it before each upgrade. Second, the cost of running a model matters little if its terms block the product. The models at the centre of the reasoning-model cost problem include MIT-licensed DeepSeek and Apache-licensed gpt-oss, which is one reason they can be self-hosted at all. Read the clause about outputs, the clause about scale, and the clause about remote restriction. Those three answers place any model in the table above.