DeepSeek

Our verdict

Open weights from one of the most-used AI products in the world, and the licence question is more interesting than the coverage suggests. "DeepSeek is MIT" is true of R1, including its weights. It is not true of V3, whose weights sit under a custom agreement carrying use-based restrictions - including a flat ban on military use - which you must pass on to anyone you distribute derivatives to. Check which model you build on before assuming MIT.

Key features

  • Open weights published for download rather than gated behind a request form
  • DeepSeek-R1, a reasoning-focused model series
  • DeepSeek-V3, a general-purpose model series
  • Code and model weights licensed separately, with different terms by family
  • Runs locally through the usual tooling, including Ollama and LM Studio
  • A hosted API offered separately from the open weights

Strengths

  • DeepSeek-R1 is unambiguous: "This code repository and the model weights are licensed under the MIT License", and the README states the series supports commercial use
  • Weights are genuinely downloadable rather than available on request
  • Under the V3 model agreement, DeepSeek "claims no rights in the Output You generate using the Model"
  • Both licence texts are published in the repositories rather than linked away
  • Open weights remove the per-token meter entirely if you have the hardware

Limitations

  • The licence is not one licence: R1 is MIT, while V3 splits code (MIT) from model weights (the DeepSeek License Agreement)
  • The V3 model agreement carries use-based restrictions in its Attachment A, including a flat prohibition on "military use in any way"
  • Derivatives of V3 must carry those same use-based restrictions forward, and recipients must be given a copy of the licence
  • DeepSeek "reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License"
  • Running either family well needs substantial GPU capacity, which is the real cost

Who should skip DeepSeek

Anyone wanting a single blanket answer on licensing across the whole DeepSeek range, since terms differ by model family. Also teams without the hardware to run large models, for whom open weights are not the saving they appear to be.

Pricing details

Checked on the official pricing page on . Prices change often, so confirm before you buy.

What the free plan includes

  • The weights are published and free to download for both model families, so there is no allowance to run out of. Your hardware is the limit, and these are large models. What differs between the families is not cost but licence terms.

Worth knowing

The open weights have no price: they are published for download and run on your own hardware. DeepSeek also sells hosted API access, and we were not able to reach its official pricing page at the time of this check, so no API figures are published here rather than repeat third-party numbers. Treat the zero price as applying to self-hosting the open weights only. Checked 23 October 2026.

Commercial use

Allowed with conditions. Two different licences apply depending on which model you take, and conflating them is the common mistake. DeepSeek-R1 ships a plain MIT License covering both code and weights; its README states that "This code repository and the model weights are licensed under the MIT License" and that the R1 series supports commercial use. DeepSeek-V3 is split: the code is MIT under LICENSE-CODE, while the weights fall under the DeepSeek License Agreement, Version 1.0, dated 23 October 2023. That agreement grants a perpetual, worldwide, royalty-free licence and expressly permits hosting the model as a service, but attaches use-based restrictions set out in its Attachment A, prohibiting among other things any use violating applicable law, "military use in any way", exploiting or harming minors, and generating verifiably false information intended to harm others. Those restrictions must be carried forward into any licence you grant over derivatives, and recipients must receive a copy of the licence. On output the agreement is generous: DeepSeek "claims no rights in the Output You generate using the Model". It also reserves the right to "restrict (remotely or otherwise) usage of the Model in violation of this License". Read from LICENSE and LICENSE-MODEL in DeepSeek's own repositories. Checked 23 October 2026.

Notable integrations

  • Ollama and LM Studio for local running
  • Hugging Face for weight distribution
  • Standard inference stacks that accept open weights
  • A separate hosted API from DeepSeek

Alternatives to DeepSeek

  • Ollama

    A local model runtime. It downloads, manages and serves open-weight language models on your own machine, and exposes them over an HTTP API on localhost.

  • LM Studio

    A desktop app that is both model runtime and chat interface. It finds and downloads models, runs them locally, and can serve them to other software over an OpenAI-compatible API.

  • Open WebUI

    A self-hosted web interface for AI models. It runs in your browser, supports multiple users with roles and permissions, and connects to Ollama or any OpenAI-compatible endpoint.

Guides