Open-weight licenses compared: Llama vs Apache, MIT and the new revenue-gated licenses

"Open model" covers very different legal terms. Some open-weight models come under textbook open-source licenses: you can build a product on them, sell it, and owe nothing but a notice file. Others are under custom licenses that cap your user count, bind you to a use policy, tell you what to name your fine-tune, or exclude whole regions. A newer group switches off once your company earns too much. We read the actual license texts on October 8, 2026. This piece explains what each one lets you do, where the traps are, and why none of them makes a model "open source" in the strict sense. It started with a practical question: when DeepSeek shipped DeepSeek-V4.1-Flash under MIT, how different is that from Llama, Gemma or Qwen?

This analysis summarizes license texts; it is not legal advice. Licenses change, and some terms here are ambiguous. Read the full text, and ask counsel before you ship.

The comparison

Last verified: October 8, 2026. Underlying rows and links are in our data file O2-licenses.csv.

MAU: monthly active users. MaaS: model-as-a-service.

Apache 2.0 and MIT: the easy cases

Apache 2.0 has become the default for mid-size open models from US and European labs. OpenAI's gpt-oss-120b and gpt-oss-20b use it. So do Google's Gemma 4 family, Mistral's Small 4, Large 3 and Ministral 3 models, according to Mistral's model docs, and Alibaba's Qwen3.8-27B. The license lets you use, modify and redistribute for any purpose. Redistribution has four conditions: include the license, mark files you changed, keep existing notices, and pass on any NOTICE file. Section 3 is the clause that sets it apart: each contributor grants you a patent license, which ends only if you sue claiming the work infringes a patent. Apache does not grant trademark rights, so you can't call your fine-tune "Gemma Pro" on the license's authority.

OpenAI adds a two-sentence usage policy to gpt-oss, which essentially asks you to comply with applicable law. Google's case is less clear. The Gemma Terms of Use, last modified April 1, 2026, state that they do not apply to Gemma 4 and point to an Apache 2.0 page. That page links to a separate "Prohibited use" page, but neither the Apache text nor Google's Gemma 4 license page says whether that policy binds Gemma 4 users (re-checked October 9, 2026). Check before relying on it.

MIT is even shorter. DeepSeek's V4.1-Flash license file is standard MIT. You may do anything with it, provided the copyright and permission notice travel with "all copies or substantial portions". The MIT text has no patent clause, which some corporate legal teams treat as a gap compared with Apache. DeepSeek's terms have changed over time: the DeepSeek-V3 repository from December 2024 split an MIT code license from a separate "Model License" for the weights. Check the exact repository you download, not the company's reputation.

Llama: permissive in practice, conditional on paper

Meta's Llama 4 Community License, effective April 5, 2025, grants a "non-exclusive, worldwide, non-transferable and royalty-free limited license". Most startups can build on it freely. The conditions are what set it apart:

• The 700M-user clause. If you or your affiliates had more than 700 million monthly active users in the calendar month before the Llama 4 release date, you must request a license from Meta, which it can grant or refuse. This affects a handful of companies, mostly Meta's direct competitors.
• Branding. If you distribute Llama or a product containing it, you must "prominently display 'Built with Llama'" and include a copy of the agreement.
• Naming. If you distribute a model built on Llama, its name must begin with "Llama". Your fine-tune can't be called "AcmeChat"; it has to be something like "Llama-AcmeChat".
• Notice. You must include a notice file stating that Llama 4 is licensed under the Llama 4 Community License, copyright Meta Platforms.
• Acceptable Use Policy. The AUP is incorporated into the license. It bars illegal uses, weapons, critical infrastructure, deception and fraud, and failing to disclose known risks to end users, among others. Breaching it breaches the license, and Meta may terminate.
• The EU clause. The AUP says that for "any multimodal models included in Llama 4", the Section 1(a) rights are not granted to individuals domiciled in, or companies with their principal place of business in, the EU. End users of products that incorporate these models are exempt. The Llama 3.2 AUP has the same clause for Llama 3.2's multimodal (Vision) models. In practice, a Berlin-based startup cannot itself take the license for image-capable Llama models, although its EU customers can use a product that a non-EU company built with them.
• Litigation and patents. If you sue Meta alleging that Llama infringes your IP, your license terminates. The grant covers Meta's "intellectual property or other rights" and never mentions patents explicitly. Unlike Apache, there is no express patent grant.
• Law. California law governs the license.

On Hugging Face, Meta's latest Llama releases are still Llama 4 Scout and Maverick, last updated May 22, 2025, so these terms remain current for Llama.

The new middle tier: revenue gates

The most notable trend of 2026 is permissive-looking licenses that switch off for large companies.

• Qwen3.8-Max License. Alibaba's 2.4-trillion-parameter flagship, Qwen3.8-2.4T-A95B, uses an MIT-style grant with two additions. First, commercial products with more than 100 million MAU or more than US$20 million monthly revenue must display the model name in their interface. Second, if you run a "Model as a Service" business, meaning you offer third parties inference or fine-tuning, or an "AI Work Assistant", meaning a standalone AI coding or office-productivity product, and your group revenue exceeds US$50 million in any 12-month period, you need a separate license from Qwen before commercial use. Internal use that doesn't expose the model to third parties is exempt. The smaller Qwen3.8-27B stays on Apache 2.0.
• Mistral's "Modified MIT". Mistral Medium 3.5 states that you are "not authorized to exercise any rights under this license" if your company's, or your employer's, global consolidated monthly revenue exceeded $20 million in the previous month. The restriction covers derivatives and combined works. Above the cap, the options are a commercial license, which Mistral may grant "at its sole discretion", or Mistral's hosted services. The license text has a minor drafting error: it refers to a clause "(b)" that is numbered "2". Mistral had not named a license for its newly announced Large 4 as of October 8; its weights aren't out yet.

These licenses aim at the businesses most likely to compete with the lab's own API: cloud resellers and coding-assistant vendors. If you are a small company, they behave like MIT. If you plan to grow past the thresholds, or to be acquired by a company that already has, they are a material risk.

"Open weights" is not "open source"

The Open Source Initiative's Open Source AI Definition 1.0, published October 28, 2024, requires the freedom to use, study, modify and share an AI system for any purpose. It also requires access to three things: the parameters (weights), the complete code used to train and run the system, and "sufficiently detailed information about the data" so that a skilled person could build a substantially equivalent system. It does not require releasing the training data itself, but it does require listing public datasets and where to get them.

By that standard, a Llama, Qwen, DeepSeek, Mistral or Gemma release that ships weights and inference code without training code and data information does not qualify, even under Apache or MIT. The license is open source; the release is not. Llama fails on the license as well. The OSI said in July 2023 that Meta's license discriminates against users (the user cap) and fields of endeavor (the AUP), both of which the Open Source Definition prohibits.

The distinction matters in law, too. The EU AI Act exempts providers of models released under a "free and open-source licence" from some documentation duties (Article 53(2)). According to WilmerHale's summary of the Commission's guidelines, the license must allow unrestricted access, use, modification and distribution, and the provider must not monetize the model. In our reading, a user cap, a binding use policy or a revenue gate sits uneasily with "unrestricted", although the Commission has not ruled on specific licenses. The exemption never applies to systemic-risk models. Our EU AI Act explainer covers the wider rules.

What this means in practice

• Building a SaaS product on a self-hosted model: Apache 2.0 or MIT models carry the least risk. Llama is fine below 700M MAU if you can live with the branding, the naming rule and the AUP. Check the revenue gates on Qwen's flagship and Mistral Medium 3.5.
• Publishing a fine-tune: under Llama, the name must start with "Llama" and you must ship the notice and agreement. Under Apache, mark changed files and keep the NOTICE. Under Gemma 1–3 terms, you must pass the use restrictions on to your users in an enforceable agreement.
• EU-headquartered companies: avoid taking the license for Llama's multimodal models directly. The Llama 3.2 clause covers only that release's multimodal models, so its text-only models are outside it. For any other Llama version, check that version's AUP.
• Patent-sensitive enterprises: prefer Apache 2.0, which has the only explicit patent grant in this set.
• Hosting models for others: read the Qwen3.8-Max "Model as a Service" definition carefully. Simply relaying requests to other hosted models is excluded.

Licensing is only one part of the open-versus-API decision; hardware and operations often cost more. See frontier APIs vs open models: what they really cost, and when you're ready to try one, run an open model locally with Ollama.