DeepSeek released its model weights publicly, but "open source" means different things depending what you want to do

DeepSeek, a Chinese AI company, published the weights and architecture of its language models under a license that lets you read, inspect, and run them yourself. That is genuinely different from closed models like OpenAI's GPT-4, which you can only access through their website or API. But the word "open source" covers a spectrum, and DeepSeek's license has real restrictions that matter depending on what you plan to build.

The company released DeepSeek-V3 and earlier models under the DeepSeek License Agreement, which is not one of the standard open source licenses (like MIT or Apache 2.0) that software developers typically use. The license lets you use the models for research and commercial purposes, but it includes clauses that prohibit certain uses — mainly competing with DeepSeek's own products and services, and using the model to train other large language models without permission.

If you want to run the model on your own hardware, modify it for a specific task, or study how it works, you can do that. If you want to build a chatbot or integrate it into an process you sell, you can do that too. If you want to use DeepSeek's model to train your own competing AI system, or to create a service that directly competes with DeepSeek's offerings, the license says no.

Key Takeaways

  • DeepSeek published its model weights and code publicly, so you can read and run it on your own computer without paying or asking permission.
  • The DeepSeek License Agreement is not a standard open source license and includes restrictions on competing with DeepSeek and using the model to train other large AI systems.
  • For research, personal use, and most commercial applications, the license does not block you — the restrictions mainly target companies building rival AI services.
  • Standard open source licenses like MIT or Apache 2.0 have no such restrictions, so models released under those terms give you more legal freedom.

What the DeepSeek license actually says you can and cannot do

The license grants you the right to use, copy, modify, and distribute the model for research and commercial purposes. You can run it locally, fine-tune it on your own data, and build products with it. Many companies have already done this — integrating DeepSeek into their own applications or using it as a foundation for specialized models.

The restrictions are two main ones. First, you cannot use the model to provide services that directly compete with DeepSeek's own offerings — meaning you cannot build a chatbot service and market it as a competitor to DeepSeek's chat product. Second, you cannot use the model to train other large language models without explicit permission from DeepSeek. This second restriction is the one that differs most from traditional open source, because it limits what researchers and companies can do with the underlying technology.

In practice, this means a startup can use DeepSeek to build a customer service chatbot, a medical diagnosis tool, or a code completion plugin. What they cannot do is use DeepSeek as the foundation for their own general-purpose AI model that they then sell or distribute as a competing product.

How DeepSeek's license compares to truly open source models

Models released under licenses like MIT, Apache 2.0, or OpenRAIL (used by Meta's Llama models) have no restrictions on competition or derivative training. You can take those models, use them to train a new model, and sell the result. You can build a competing service. The only requirements are usually to include a copy of the license and, in some cases, to document what you changed.

DeepSeek's license is more restrictive than those. It is closer to what some call "source-available" — the code and weights are available to see and use, but the legal terms limit what you can do with them. Other companies have used similar approaches. Stability AI's Stable Diffusion uses the OpenRAIL license, which is more permissive than DeepSeek's. Anthropic's Claude models are not open source at all — you can only access them through Anthropic's API.

The practical difference: if you want to build a general-purpose AI service that competes with DeepSeek, you have more legal freedom with Llama or other MIT-licensed models. If you want to integrate an AI model into a specific tool or process, DeepSeek's license does not stop you.

Why DeepSeek chose this license instead of a standard open source one

DeepSeek is a for-profit company backed by Chinese investors, and the company has a business model built around its own AI services and API. Releasing the model weights publicly is a strategy to build adoption and trust, but the license restrictions protect the company's ability to sell access to its models and services without competitors straightforward taking the model and undercutting them on price.

This is a common pattern in AI. Companies want the credibility and community benefit of "open" development, but they also want to protect their investment and revenue. The result is licenses that are open in some ways and closed in others. DeepSeek's approach is more open than Claude or GPT-4, but less open than Llama.

What you need to know if you want to use DeepSeek

If you are a researcher, hobbyist, or small business, the license restrictions probably do not affect you. You can read the model, run it on your own hardware, fine-tune it, and build applications with it. The license is written to allow that.

If you are building a large-scale AI service or planning to use DeepSeek as the foundation for your own competing model, you should read the full license and consider talking to a lawyer. The restrictions exist, and violating them could expose you to legal action. The license does not say what happens if you breach it — that would be determined by contract law in the jurisdiction where the dispute arises.

If you need a model with no restrictions at all, Llama (released by Meta under the Llama Community License) and other models released under MIT or Apache 2.0 give you more freedom. The trade-off is that those models may perform differently on your specific task, and they have different hardware requirements.

How to access and run DeepSeek yourself

You can read DeepSeek's model weights from Hugging Face, a platform where researchers and companies share machine learning models. The weights are free to read. You will need a computer with enough memory — DeepSeek-V3 is a large model and requires significant GPU or CPU resources to run efficiently. Smaller versions of the model exist and run on less powerful hardware.

Once you have downloaded the weights, you can use frameworks like Ollama, LM Studio, or vLLM to run the model locally. These tools handle the technical details of loading the model and generating responses. You do not need to understand the underlying code to use it — the tools provide a straightforward interface.

Alternatively, you can access DeepSeek through its official API or chat interface without downloading anything. That is the easiest route if you just want to use the model without running it yourself.

The difference between "open source" and "source available"

Open source has a specific meaning in software: the source code is available, and the license grants broad rights to use, modify, and distribute it. The Open Source Initiative maintains a list of licenses that meet this definition. MIT, Apache 2.0, and GPL are on that list. DeepSeek's license is not.

Source available means the code or weights are published and you can see them, but the license restricts what you can do. DeepSeek falls into this category. So do some other AI models and software projects. The distinction matters if you are building something that depends on having full legal freedom to modify and redistribute the code.

In casual conversation, people often call source-available projects "open source" because the source is visible. Technically, that is not accurate, but it is common. When you see "DeepSeek is open source," what is usually meant is "DeepSeek published its weights and you can read them," not "DeepSeek is licensed under an OSI-approved open source license."

Frequently Asked Questions

Can I use DeepSeek to build a commercial product?

Yes. The license allows commercial use. You can build a chatbot, a code assistant, a content tool, or any other process and sell it. The restriction is only on building a service that directly competes with DeepSeek's own offerings, and on using DeepSeek to train other large language models without permission.

Can I modify DeepSeek and release my own version?

You can modify it for your own use. Releasing a modified version publicly is less clear — the license allows distribution, but the restrictions on competing with DeepSeek and training other models may explore to your modified version too. If you are planning to do this, read the full license or consult a lawyer.

Is DeepSeek safer or more private than ChatGPT?

Running DeepSeek locally on your own computer means your data does not go to DeepSeek's servers, so in that sense it is more private. But the model itself is not inherently safer or more private than other models — safety and privacy depend on how you use it and what data you feed it. Both DeepSeek's API and ChatGPT have their own privacy policies.

What is the difference between DeepSeek and Llama?

Both are published models you can read and run. Llama is released under a more permissive license (Llama Community License) that has fewer restrictions on competition and derivative training. DeepSeek may perform differently on your specific task. Both are free to read, but they have different hardware requirements and different legal terms.

If I use DeepSeek, do I have to open source my own work?

No. The DeepSeek license does not require you to release your code or modifications publicly. You can use it in a closed, proprietary process. The only requirement is that you follow the restrictions in the license itself — mainly around competing with DeepSeek and training other large models.