Chinese artificial intelligence firm DeepSeek on Thursday unveiled V3.1, an upgrade to its flagship model designed to deliver faster processing and—critically—compatibility with China’s upcoming generation of domestically produced chips.
The move underscores Beijing’s determination to build a self-reliant semiconductor ecosystem in response to U.S. export restrictions that have limited access to advanced Western technologies.
DeepSeek, which jolted the global AI industry earlier this year by releasing cost-effective models rivaling Western systems like OpenAI’s ChatGPT, said V3.1 features UE8M0 FP8 precision optimization, enabling more efficient use of memory while increasing speed. FP8—or 8-bit floating point—is widely seen as the next step in reducing computing costs for large-scale AI applications.
While the company did not reveal which Chinese-made chips would be supported, it hinted that the model had been tuned for “soon-to-be-released next-generation domestic chips.”
V3.1 also introduces a hybrid inference structure that lets users toggle between reasoning and non-reasoning modes via a new “deep thinking” button on DeepSeek’s app and web platform. Developers using its API will see a revised cost structure beginning September 6.
The update comes after two other significant releases this year: an R1 model update in May and an earlier V3 enhancement in March. Analysts say the focus on local chip support positions DeepSeek as a key player in China’s long-term plan to reduce reliance on foreign AI infrastructure.
