Is DeepSeek open source? Yes, weights are MIT, data is not

DeepSeek's current models are open weight, and most of them carry the MIT license. That is the short answer to whether DeepSeek is open source. DeepSeek's Hugging Face cards for V4.1 Flash, V4.1 Flash 1017 and V4 Pro all say the repository and the weights are licensed under MIT, which allows commercial use. What DeepSeek has not published is the training data or the training code, so "open source" in the strict sense does not fit. "Open weights" does.

The distinction matters if you plan to build on these models. MIT weights mean you can download, run, fine-tune and ship them. They do not let you reproduce how the model was made.

Which DeepSeek models have MIT weights

Here is what the Hugging Face cards showed on November 1, 2026.

Treat the last two rows with care. Each model has its own repository, and the license sits in that repository, not in a blanket statement for the whole company. Check the card for the exact model you download.

DeepSeek's GitHub also holds code. Its list of repositories includes GPU kernel libraries such as FlashMLA and DeepGEMM, and a deepseek-harness framework, all labeled MIT, plus an inference framework called DeepInfer labeled Apache-2.0. Those labels come from the repository list, and we did not open each license file.

What DeepSeek has not released

The V4.1 Flash card says the model was trained from scratch on a multimodal corpus of 45 trillion tokens. It does not say what is in that corpus. The training data is not published. The card links a technical report and includes folders for inference and prompt encoding, but we found no training code for the model. The Hugging Face page lists the benchmark scores, which come from DeepSeek's own evaluation setup, not from independent tests.

So you can inspect the weights and run them, but you can't audit what they were trained on or retrain them from scratch.

API or self-hosting

You can use DeepSeek two ways.

The API. DeepSeek's API docs list two model names, deepseek-flash and deepseek-v4-pro. The API follows the OpenAI format at https://api.deepseek.com, and there is an Anthropic-compatible endpoint at https://api.deepseek.com/anthropic. It is billed per token. As of November 1, 2026, the pricing page lists deepseek-flash at $0.15 per million input tokens on a cache miss and $0.60 per million output tokens off-peak. deepseek-v4-pro costs $0.66 and $1.98. Peak rates are double. Both models have a 1M-token context. See our guide on whether DeepSeek is free for the full cost picture.

Self-hosting. You download the weights from Hugging Face and run them yourself. The scale is the catch. V4.1 Flash has 763B parameters on its listing, and V4 Pro has 1.6T. These are models for server clusters, not laptops. Your data stays on your own hardware, and you pick the infrastructure. For a closer look at the Flash model, read our explainer on DeepSeek V4.1 Flash.

The two routes differ on data. Using DeepSeek's own app or API sends your prompts to DeepSeek. Self-hosting does not.

Is DeepSeek free, safe and Chinese

The weights are free to download under MIT, but running them costs hardware. The API charges per token, and DeepSeek's site presents its chat as free to try.

On safety, the question is mostly about data. DeepSeek's privacy policy says it collects, processes and stores personal data in the People's Republic of China, and that it collects prompts, uploaded files and chat history. Self-hosted weights avoid that. We did not test the models for safety or security in this article.

DeepSeek is Chinese. The service is run by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., according to its privacy policy.

What to check before you build on it

Read the license in the repository of the model you pick, since cards differ. Then decide whether you need the API, where your data goes, or your own servers. Prices and model names have changed several times this year, so check the pricing page again before you commit.