DeepSeek Releases V4-Pro, an MIT-Licensed MoE Model

DeepSeek has released DeepSeek-V4-Pro, the latest entry in its widely watched line of open models, publishing the weights on Hugging Face. The model is a mixture-of-experts (MoE) system shipped with FP8 weights and made available under the permissive MIT license.

The release is positioned as a text and reasoning model, continuing DeepSeek's push into systems that can handle both general language tasks and multi-step problem solving. As with the company's earlier work, distributing the model in FP8 format signals an emphasis on efficient inference, letting the large expert-based architecture run with a smaller memory footprint than full-precision equivalents.

Why it matters

DeepSeek has become one of the most closely tracked labs in open-weight AI, and each major version tends to reset expectations for what freely available models can do. A few points stand out here:

• MIT licensing places few restrictions on commercial and research use, a key draw for teams building on open models.
• Mixture-of-experts design activates only part of the network per token, a route to scaling capacity without proportional compute cost.
• FP8 weights lower the barrier to deployment on constrained hardware.

Detailed specifications such as parameter counts, context length, and benchmark results were not included in the release record, so builders will want to consult the official model card as DeepSeek fills in documentation. For now, the arrival of a new MIT-licensed flagship gives the open-source community another capable foundation to test, fine-tune, and deploy.