DeepSeek Releases V4-Pro, an Open MoE Contender

AI research company DeepSeek has released DeepSeek-V4-Pro, a significant new open-source model aimed at high-performance reasoning, coding, and text generation. The model enters a competitive field with a compelling combination of advanced architecture and an unrestrictive license.

Technically, DeepSeek-V4-Pro is a Mixture-of-Experts (MoE) model, an architecture known for its computational efficiency during inference. The release also utilizes fp8 precision for its weights, a format that can further improve performance and reduce memory usage without a significant loss in accuracy, making it more accessible for developers to run on consumer-grade hardware.

Why the License Matters

What truly sets DeepSeek-V4-Pro apart in the current landscape is its license. The model is released under the MIT License, one of the most permissive open-source licenses available. This allows for unrestricted use, modification, and distribution, including for commercial applications, without the copyleft provisions or use-case restrictions seen in other popular model releases. This choice signals a strong commitment to fostering a truly open and commercially viable ecosystem around the model.

This combination of a powerful MoE architecture and a fully permissive license makes DeepSeek-V4-Pro a noteworthy option for startups and established companies alike looking to build on open foundations. The model is available now on Hugging Face for researchers and developers to explore.