The update, known as DeepSeek-V3.1, introduces a feature that optimizes the model for compatibility with upcoming Chinese-made chips, a move that comes as Beijing accelerates efforts to reduce reliance on U.S. technology in the wake of Washington’s export restrictions. In addition, the model promises faster processing speeds and improved efficiency.
DeepSeek, which captured global attention earlier this year by launching AI systems capable of competing with Western leaders such as OpenAI while offering lower operational costs, has been updating its suite of models at a rapid pace. The new release follows the rollout of its R1 model in May and a prior V3 enhancement in March, signaling the company’s determination to maintain momentum in an increasingly crowded market.
At the technical level, DeepSeek said the V3.1 model employs the UE8M0 FP8 precision format, optimized for “soon-to-be-released next-generation domestic chips.” While the company did not specify which chipmakers or models would be supported, the focus highlights the importance of the FP8 format. FP8, or 8-bit floating point, allows large AI systems to run more efficiently by consuming less memory while delivering faster performance.
The V3.1 model also introduces a hybrid inference structure, enabling it to switch between reasoning and non-reasoning modes. Users can toggle this function through a “deep thinking” button available on DeepSeek’s official app and web platform, both of which have now migrated to the upgraded version.
Beyond technical upgrades, the company revealed plans to adjust its API pricing model—which allows developers to embed DeepSeek’s AI into third-party applications—starting September 6.
Taken together, the enhancements reflect DeepSeek’s dual strategy: improving AI efficiency for end users while ensuring its models remain in sync with China’s broader push to cultivate a self-reliant technology ecosystem.