DeepSeek's Model Licence: What the Terms Permit

“DeepSeek is MIT licensed” is true of some checkpoints and false of others, and it is never true of the whole repository in one sentence. The licence you are under depends on which checkpoint you downloaded and which file you are asking about.

Two licences per repository

DeepSeek’s open releases have consistently separated two things that people conflate:

• The code — the inference scripts, the tooling, the repository itself — released under the MIT Licence. Permissive, short, no use restrictions.
• The weights — the model itself — governed by a separate model licence, historically the DeepSeek Model Licence, which permits commercial use but attaches a schedule of use restrictions.

A GitHub repository showing an MIT badge therefore tells you nothing about your right to serve the model, because the badge is describing the code. The weights are usually distributed from Hugging Face and carry their own terms, and in the DeepSeek repositories those terms have lived in a separate file — commonly a LICENSE-MODEL alongside a plain LICENSE for the code.

R1 and the move to MIT

With the DeepSeek-R1 release in January 2025, DeepSeek stated that the R1 series is made available under the MIT Licence and that this explicitly permits commercial use, modification, derivative works, and distillation — that is, using R1’s outputs to train other models. That last permission is unusual and is the clause that distinguishes this release from most open-weight terms, which commonly forbid using outputs to improve a competing model.

DeepSeek-V3’s weights were initially published under the DeepSeek Model Licence rather than MIT, and a later V3 checkpoint was published under MIT. This is why the question “what licence is DeepSeek under” has no single answer: the terms are attached to checkpoints, and they changed between them.

The distilled models inherit

This is where most real licence problems occur, and it is entirely predictable once stated. The R1-distill models are not DeepSeek architectures. They are Qwen and Llama base models fine-tuned on reasoning data generated by R1. A fine-tune is a derivative work of its base, so the base model’s licence continues to apply to the result.

What that means in practice

• The Qwen-based distills derive from Qwen2.5 models, which Alibaba released under the Apache 2.0 licence for the sizes used here. Apache 2.0 is permissive and imposes attribution and notice obligations rather than use restrictions.
• The Llama-based distills derive from Llama 3.1 and Llama 3.3, which are released under Meta’s community licence rather than an OSI-approved open-source licence. That licence carries an acceptable-use policy, a naming requirement for derivative models, and a threshold above which separate permission is needed.
• The DeepSeek contribution — the fine-tuning, the data, the released weights — is governed by DeepSeek’s terms for that repository. You are under both sets at once.

The consequence worth writing down: two files sitting in the same directory on your disk, both named for the same R1 release, can be under materially different obligations depending on which base they came from. The page on distill context windows lists which base sits behind which size, and it is the same table you need here.

Use restrictions in the DeepSeek licence

Where the DeepSeek Model Licence applies rather than MIT, it follows a pattern common to several open-weight releases of that period: a broad grant including commercial use, subject to an attached schedule of prohibited uses, with an obligation to pass the same restrictions on to anyone you distribute the model or a derivative to.

The prohibited-use schedules in this family of licences cover categories such as unlawful use, military application, harm to minors, generating false information presented as genuine, and uses that discriminate against individuals or groups. The specific enumeration is the operative text and it is short; read it rather than a summary, including this one.

The structural point for engineering teams is that these are downstream-propagating obligations. If you ship a product that embeds the weights, or distribute a fine-tune, the restrictions travel with what you shipped and you are expected to carry the terms forward. A permissive licence like MIT or Apache 2.0 has no equivalent mechanism, which is the practical difference between the two regimes.

How to check the one you have

1. Identify the exact repository and revision. Not the model family — the repository path and the commit hash you pulled. Terms attach to that.
2. Read both licence files. A repository with a LICENSE and a LICENSE-MODEL is telling you the code and the weights differ. A repository with only one may still be constrained by its base model.
3. Identify the base model. For anything with Distill in the name, the base is named in the model card and its licence applies too. Read that one as well.
4. Record the answer with your artefact. Store the repository, the revision and the licence text alongside whatever you built. Reconstructing which terms applied eighteen months later, after the upstream repository has been relicensed, is very difficult.
5. Separate the licence from the hosted terms. The weights licence governs weights you run. Calling DeepSeek’s API is governed by its terms of service instead, which are a different document with different obligations — the usage policy for the hosted API covers that side.

Questions a weights licence does not answer

Teams often go to the licence for answers it was never written to give, and then treat silence as permission. Four questions come up repeatedly and none of them is fully settled by the licence text.

• Who owns the output. A model licence governs the model. What you may do with generated text is addressed — if at all — in different terms, and the underlying legal position on machine output varies by jurisdiction and is not uniform. If ownership of output matters commercially, that is a question for counsel and not for a LICENSE file.
• Whether the training data was licensed. A permissive licence on weights says what the publisher grants you. It is not a warranty about the provenance of what the model was trained on, and open-weight releases generally disclaim warranties broadly. This is the substance behind procurement questions about indemnity, which open weights typically do not carry and some commercial APIs do.
• Whether you may train on the outputs. Some open-weight licences forbid using outputs to improve another model, some are silent, and the R1 announcement is notable precisely because it permits distillation explicitly. Silence is not consent here — check the specific text, and check the terms of any hosted API whose outputs you used.
• What you owe your own users. None of these documents imposes disclosure, safety or accessibility obligations on your product. Those come from your regulator, your customers and your own policy, and a permissive licence removes no obligation you would otherwise have.

The practical division is that a licence tells you what you may do with the artefact and almost nothing about what you should do with the system you build from it. Answer the first from the file; answer the second somewhere else.