DeepSeek releases downloadable model weights, and many of its major models use the MIT license, which permits commercial use. That includes R1, R1-0528 and the V4 Pro and Flash releases. But the answer changes with the exact download: the original V3 has a separate model license, and R1 models built on Qwen or Llama also require attention to their underlying licenses.
For a developer, the useful question is what you can download, modify and put into a product. For someone using the chatbot, the more immediate question is where their messages go. An open model license answers the first question; it does not determine the privacy rules of a hosted app.
This guide explains those differences using the published model cards and license texts. If you are looking for an overview of the product rather than its licensing, start with our DeepSeek guide.
On this page: What is open · Licenses by model · Commercial use · Check a download · Hosted vs local · Safety and privacy · FAQ
What does “open source” mean for DeepSeek?
There are three different things to look at: the trained model, the software that runs it, and the information needed to understand how it was trained. A release can be generous in one area without providing everything in the others.
Model weights are the learned numerical parameters used to generate responses. Downloadable weights let you run a model on infrastructure you choose, provided you have compatible software and enough hardware. The attached license determines your permission to modify or redistribute it. Public availability alone is not a license.
Inference code runs an already trained model. A repository may contain this code, chat templates and setup instructions without containing the full training pipeline. A technical paper adds explanations, but reading a paper is different from receiving everything needed to reproduce the system.
The Open Source Initiative’s Open Source AI Definition 1.0 goes beyond downloadable weights. It calls for the freedom to use, study, modify and share a system, with the preferred form for modification including parameters, code and sufficiently detailed data information. That includes training and data-processing code, rather than only a script that loads the finished model.
One detail often gets lost in this debate: OSI does not simply require every raw training example to be released. Its explanation of the data requirements distinguishes shareable data from information that cannot legally be redistributed. Detailed information about the data still matters.
For DeepSeek, “downloadable weights under the license listed for this checkpoint” is the most useful precise description. An MIT license on released files does not, by itself, establish that an entire AI system meets the OSI definition. Equally, the terminology debate does not erase the practical freedom those downloads give you: you can choose your hosting, retain a particular version and build your own application around it.
DeepSeek licenses by model and version
Start with the full repository name. “DeepSeek,” “R1 8B” or “V3” can hide a distinction that changes the terms you need to read. The following table covers the major releases and the exceptions most likely to affect a download or commercial project.
| Release | Published licensing | What to watch |
|---|---|---|
| DeepSeek-R1 | MIT for the code repository and model weights. | The card expressly supports commercial use, modifications and derivatives, including distillation. |
| DeepSeek-R1-0528 | MIT; commercial use and distillation are supported. | This is the full model. Do not confuse it with the separate Qwen3-8B release. |
| R1-Distill-Qwen 1.5B, 7B, 14B and 32B | DeepSeek’s R1 release uses MIT; the Qwen2.5 bases are identified as Apache 2.0 in the R1 license notes. | Keep the applicable upstream license and notices with the derivative. |
| R1-Distill-Llama-8B and R1-Distill-Llama-70B | DeepSeek’s R1 terms plus the relevant Llama base-model terms. | The bases are Llama 3.1-8B and Llama 3.3-70B-Instruct respectively. These are not MIT-only downloads. |
| R1-0528-Qwen3-8B | DeepSeek publishes it under MIT; its Qwen3 base has Apache 2.0 licensing. | It is a different model from R1-Distill-Llama-8B despite sharing the 8B size label. |
| Original DeepSeek-V3 | Code: MIT. Base and Chat model weights: DeepSeek Model License. | Commercial use is supported, but the model license contains use-based and redistribution conditions. |
| DeepSeek-V3-0324 and DeepSeek-V3.2 | MIT for their repositories and model weights. | Do not apply the original V3 license description to these later releases. |
| DeepSeek-V4-Pro, DeepSeek-V4-Flash and DeepSeek-V4.1-Flash | MIT. | These downloadable releases are separate from the terms and charges of a hosted service using them. |
| V4-Pro-0813 and V4-Flash-0731 | MIT for the repositories and model weights. | The dated releases have their own download pages. Their cards confirm the license for those specific checkpoints. |
The original V3 versus V3-0324 distinction is especially important. A statement that “V3 uses DeepSeek’s custom license” is accurate for the original release but misleading when applied to the whole family. Checkpoint names are part of the licensing decision, not just labels for performance updates.
The same applies to small R1 models. A distilled model learns from a larger model’s responses while using another model as its starting point. That underlying model can bring its own obligations. Our R1 model guide explains the family and deployment choices; the table above addresses permission to use the specific releases.
Can you use DeepSeek commercially?
Yes, the releases listed above support commercial use under their applicable terms. What changes is the set of conditions attached to that permission. A paid customer-support app, a redistributed model download and an internal research tool are different uses, so check the parts of the license that match what you plan to do.
What MIT allows—and what you still need to keep
The MIT license permits use, modification, distribution and sale of the licensed material. Its notice condition requires retaining the copyright and permission notices in copies or substantial portions. It also supplies the material without warranty.
MIT does not require you to publish the source code of your entire application simply because it uses an MIT-licensed model. For example, a paid writing app can keep its own application code private while complying with the model’s notice requirements. You still need to account for any other components included in the product.
A practical way to handle this is to keep the downloaded license alongside the model in your project and include the required notices in any package you distribute. Do not replace the copyright notice with your own company name. This is a summary of the published terms; the actual license is the reference for a distribution decision.
Permission to use a model is also different from permission to present your product as official or endorsed. Describe your own application honestly, and review any applicable brand terms separately. Likewise, a model license is not a guarantee that every generated passage is accurate or clear of third-party rights.
Why Qwen and Llama derivatives need a second check
For the Qwen-based releases above, Apache 2.0 is also permissive, but it has its own redistribution conditions. These include supplying the license, marking modified files, retaining relevant notices and carrying applicable NOTICE attributions when provided. The Apache license text also includes an express patent grant with conditions. A DeepSeek MIT label does not mean you should discard those upstream materials.
The Llama 3.1 and Llama 3.3 community licenses have different requirements. They include attribution and acceptable-use provisions, a “Built with Llama” display requirement in specified distribution or product situations, and naming conditions for certain models you create and make available.
Both also have additional commercial terms for organizations whose products or services exceeded 700 million monthly active users in the month before the relevant Llama version’s release date. That condition is tied to the dates and scope stated in the agreements; it is not a general claim that ordinary commercial use requires a paid license.
If your project can use either a Qwen-based or Llama-based checkpoint, licensing can be a sensible part of the choice alongside hardware and output quality. Establish the base model before you invest in adapting it.
The original V3 has its own conditions
The original V3 Model License permits commercial deployment but includes restrictions on uses such as military applications and certain harmful or discriminatory activities. Redistribution requires a license copy, relevant notices and notice of modified files. Its use restrictions must also be passed on as enforceable conditions. The definition of distribution includes providing hosted access, so keeping the weights on your server does not make those provisions irrelevant.
This is why checking only a repository’s MIT-licensed code is insufficient. Read the license for the weights as well. For an unusual commercial arrangement, have the exact license and intended deployment reviewed together rather than relying on a family-wide label.
Can you fine-tune, redistribute or distill DeepSeek?
For R1, DeepSeek explicitly allows modifications and derivative works, including using it to distill another model. R1-0528’s documentation also supports commercial use and distillation. That gives you a documented starting point for adapting these releases instead of assuming that a free download is limited to personal research.
Keep three sets of permissions separate: the source model, the model you are training, and the data you use. Permission to distill R1 does not remove the license of a different base model you fine-tune. Likewise, possession of a dataset does not establish that you may train on it or redistribute it.
For illustration, imagine adapting a Qwen-based R1 checkpoint to answer questions about your company’s public manuals. The model review covers DeepSeek and the Qwen base. The data review covers whether you control those manuals and whether they contain third-party material. If you release the adapted weights, the distribution review covers the required licenses and notices. Those are related checks, but one cannot substitute for the others.
If you obtain training responses through a hosted API, its service terms matter too. DeepSeek’s Open Platform Terms, section 4.2, expressly include training other models and distillation among permitted uses of inputs and outputs, subject to the terms and applicable law. Do not assume another provider offers identical rights just because it serves a DeepSeek model.
How to check the license of the model you download
You do not need to audit every DeepSeek repository to make a sensible choice. Follow the particular file you intend to use back to its source.
- Record the full model name and publisher. Start from the official DeepSeek model links in the table. Note the repository and revision you download. A short name in a desktop app can conceal the original checkpoint or a community conversion.
- Read the model card and license files together. Look for files such as LICENSE and LICENSE-MODEL, and check whether the terms cover code, weights or both. A badge is helpful navigation, but the actual terms and card provide the context.
- Follow the base-model reference. For a distilled, fine-tuned or merged model, identify what it was built from. Keep the relevant upstream license and attribution files rather than only the last publisher’s license.
- Match the terms to your use. Running a private copy, serving paying customers, releasing modified weights and training another model can trigger different provisions. Record the notices or restrictions that apply to your route.
- Keep the record with the deployed version. Save the license texts, source links and revision with your project. Revisit them when replacing the model, rather than assuming that an old approval covers a similarly named release.
For example, if a download says “R1 8B,” first establish whether it means R1-Distill-Llama-8B or R1-0528-Qwen3-8B. The first leads you to Llama 3.1 terms; the second leads you to the Qwen base and Apache 2.0. That one check is more useful than relying on a broad statement that all R1 downloads are MIT.
A GGUF conversion or other quantized copy makes the model easier to load in some tools; it does not make the original license disappear. If the uploader does not identify the base model or provide enough licensing information, choose a better-documented download before building around it.
Open weights, the chat app and the API are different things
Downloading a model lets you operate your own copy. Using DeepSeek’s website or API gives you access to a service operated by someone else. Both can involve the same model family, but they give you different control over the software, data and operating costs.
Official web and app: these are hosted services. A model’s MIT license does not establish that the complete chat application, account system, search integration and hosting infrastructure are available for you to copy. Use the service’s terms and privacy notice to understand that relationship.
API or third-party hosting: you send requests to the provider’s infrastructure. You pay for the service under its terms, even when the underlying weights can be downloaded without a model-license fee. DeepSeek’s Open Platform has separate service conditions and billing provisions; a different host may have different ones.
Self-hosting: you choose the model version, serving software and infrastructure. That can be a laptop for a suitable smaller model or a much larger server setup. You take responsibility for installation, access controls, updates and capacity.
“Open” therefore does not mean “free to operate.” Local hardware, electricity, storage, rented GPUs and staff time still cost money. A free model license also does not promise that a large checkpoint will fit on an ordinary computer. Check the hardware requirements and model-size choices before downloading.
For occasional questions, a hosted service is usually the simpler starting point. Self-hosting becomes more compelling when you need a fixed checkpoint, offline processing, model modifications or direct control over the system. Those benefits need to justify the work of operating it. A business handling restricted data should compare actual contracts and controls instead of treating either route as automatically compliant.
Does open source make DeepSeek safer or more private?
It gives you more ways to investigate and control a deployment. It does not guarantee accurate answers, secure software or private data handling.
The International AI Safety Report 2026 describes both sides: open weights broaden research and independent evaluation, but safeguards are easier to remove and copies cannot all be recalled after release. An update from the original developer also cannot ensure that every independent deployment installs it.
For your own application, separate answer quality from information security. A model can keep a document on your server while still giving a wrong answer about it. Conversely, a useful answer does not prove that the system stored the prompt appropriately. Downloadable weights let you evaluate and change behavior; they do not certify the result.
Where your messages go depends on the deployment
For official services covered by DeepSeek’s Privacy Policy, the policy describes collection of prompts, uploads, chat history and device or network information. It states that DeepSeek directly collects, processes and stores personal data in China. These disclosures matter when deciding what to submit to the covered services.
The policy explicitly excludes the processing rules for end users of downstream applications built using the Open Platform. Those applications need their own privacy disclosures. Check the developer’s notice, applicable API terms and any contractual commitments about retention, training use and data location; do not infer a universal API guarantee from the consumer policy.
With a properly configured local deployment, prompts and outputs can stay within your environment. But a local model connected to an external search API, cloud logging or a remote document service can still transmit information. Map the whole request path, including integrations and logs, before describing a system as private or offline.
What to check before self-hosting with sensitive data
The original model is only one component. Use these checks to turn the control offered by open weights into an actual deployment decision:
- Know what you installed. Confirm the source, checkpoint and license. Use trusted serving software and safer weight formats where supported; do not load unreviewed remote code just to get a download working.
- Control access and connections. Keep the inference service behind appropriate network restrictions and authentication. Grant users and tools only the access they need, and restrict unnecessary outbound connections.
- Decide what is retained. Include prompts, outputs, uploaded files, embeddings, logs and backups. Avoid recording sensitive content unnecessarily, and set access and deletion rules for what remains.
- Limit what an answer can do. Add appropriate input and output safeguards, test misuse scenarios before production and require review for consequential actions. Treat retrieved documents and model-generated instructions as untrusted input to other systems.
- Assign ongoing responsibility. Patch the serving software, monitor abuse without overcollecting content, and decide who responds to a leak or harmful output. Review the requirements for the data and jurisdiction before introducing regulated information.
These are practical checks, not a security certification. If no one can maintain the deployment, the extra control of self-hosting may not translate into better protection. Our DeepSeek safety guide covers the broader user-facing privacy and security decision.
Frequently asked questions
Is DeepSeek R1 fully open source?
Its released code and weights are MIT-licensed, with commercial use and adaptation permitted. Calling the entire training system fully open source is a broader claim: the MIT license alone does not establish that all the data information and code required by the OSI definition have been supplied. Also distinguish the full model from Qwen- and Llama-based derivatives.
Is DeepSeek V4 open source too?
The V4-Pro, V4-Flash and V4.1-Flash downloads listed above use MIT licensing. That gives you rights in the released material, while running them still requires suitable hardware and software. See the V4 guide for how the releases differ in use.
Does open source mean DeepSeek is completely free?
No. Downloadable MIT-licensed weights do not come with a charge per generated token from the model license, but operating them has costs. A hosted API is a separately provided service and may charge for usage. Choose between licensing costs and operating costs explicitly rather than combining them into a single “free” label.
Do I have to open-source an app that uses DeepSeek?
MIT does not impose that requirement on your application. You must still satisfy its notice condition and the terms of other components you distribute. If your model is a Llama derivative or the original V3, review those terms too; this answer is not a blanket rule for every DeepSeek-branded download.
Does downloading the weights give me the training data?
No. Weights are the trained parameters, not a copy of the original dataset. A model card, paper, dataset release and software repository are separate materials. Check what the specific project publishes rather than assuming that one model download includes all of them.
Can DeepSeek work offline?
A compatible local model and runtime can generate responses without calling a hosted model API once the required files are installed. Features that fetch live web information or contact external services still need a connection. Check the application’s integrations as well as where inference runs.
Does open source mean the model is uncensored?
No. A license describes permission to use material; it does not describe which requests a trained model will answer. Refusals, behavior changes and safety controls depend on the checkpoint and the application around it. A community label such as “uncensored” is not proof of reliability or security.
Choose the specific release before building around it
DeepSeek’s downloadable models give you a useful choice: use hosted access for convenience, or operate and adapt a copy yourself. Before committing to the second route, identify the exact checkpoint, keep its license and any upstream notices, and confirm that you can support the deployment.
For a first local project, start with a model that fits your hardware and non-sensitive sample material. Once you can explain both its licensing and where its data goes, you have a much firmer basis for deciding what to build next.