DeepSeek updates R1 AI model and releases it on Hugging Face

Beijing-based AI startup DeepSeek has rolled out an updated version of its high-profile R1 reasoning model, making it available for public access via the developer platform Hugging Face. The announcement was made Wednesday morning through a message on the company’s official WeChat account.

According to DeepSeek, the latest version of R1 is a “minor” upgrade to its original release. The Hugging Face repository currently lacks detailed documentation. However, it includes the essential configuration files and model weights—the underlying parameters that define how the AI system functions.

Notably, the model is being offered under the MIT license, a permissive open-source framework that allows commercial use without significant restrictions. This move opens the door for developers and companies worldwide to experiment with or integrate R1 into commercial applications.

Massive Scale and Hardware Demands

The updated model is enormous in scale, boasting 685 billion parameters, a number that underscores its complexity and capability. However, due to its massive size, running the model without modifications on consumer-grade hardware is unlikely. The parameter count, often synonymous with “weights,” reflects the model’s potential but also highlights its heavy resource requirements.

DeepSeek made headlines earlier this year with the debut of its original R1 model, which reportedly matched or even outperformed offerings from leading AI firms like OpenAI. This surge in visibility has also drawn scrutiny. Some U.S. regulators have raised concerns, citing national security risks associated with DeepSeek’s advanced technology and its rapid development pace.