DeepSeek launches V3.1 AI model with major upgrades

Chinese AI startup DeepSeek has unveiled V3.1, an upgraded version of its widely watched artificial intelligence model, adding new capabilities that could shake up the global AI race. The Hangzhou-based company announced the release through its official WeChat group, confirming the model is ready for testing and now available for download on Hugging Face. DeepSeek V3.1 significantly expands the model’s power, boasting 685 billion parameters and a much longer context window of 128,000 tokens. This enables it to process the equivalent of two full-length English novels in a single prompt, allowing for longer conversations and more detailed recall. The model also supports multiple tensor formats, including BF16, F8_E4M3, and F32, which broaden its usability for developers. However, the API and inference service have not yet been launched. The release follows the controversy stirred by DeepSeek’s earlier V3 model, which the company claimed cost just $5.6 million to train using about 2,000 slower Nvidia chips. That undercut the billions typically spent on training frontier models from OpenAI, Google, and Anthropic, and triggered a one-day $600 billion drop in Nvidia’s market value. Yet despite its technical promise, the chatbot faced bans in several countries due to concerns over user data being stored on Chinese servers. Early testers note that V3.1 is “very verbose,” with changes such as the removal of the r1 in the think button, suggesting it may function as a mixed reasoning model. DeepSeek’s R1 reasoning model is already available through major U.S. cloud providers—AWS, Microsoft Azure, and Google Cloud—where data is hosted locally to address security concerns. Developers are also awaiting R2, the next generation of R1. DeepSeek’s approach underscores China’s commitment to open-source AI, a strategy backed by the government’s five-year development blueprint. Unlike U.S. rivals that rely on closed, proprietary systems, Chinese companies such as DeepSeek, Alibaba, and Moonshot are releasing free-to-use models to accelerate adoption and compete globally. By lowering costs and removing access barriers, China hopes to position its AI models as viable alternatives to dominant U.S. offerings.