Tech Desk 
Aug 21, 2025
DeepSeek, a Chinese AI startup, has announced an upgrade to its flagship V3 model that includes various improvements to operation speed, as well as a feature to optimise performance on domestic chips to be released soon.
This indicated the growing emphasis on China's aim to build an independent semiconductor ecosystem, especially since Beijing is working to detach itself from US technology under expanding export restrictions and bans.
DeepSeek was already making headlines earlier in 2023 for the V3 model, as the company was competing with western stakeholders, such as OpenAI's ChatGPT, to offer competitive AI models supporting cheaper operational models.
The latest performance upgrade, labelled DeepSeek-V3.1, builds upon enhancements to the company's core models, which comes after they had released an R1 model in May and a previous upgrade to V3 in March.
In statements made through WeChat, the company detailed that the new V3.1 model supports the UE8M0 FP8 precision format, which is optimised for "the next-generation domestic chips soon to be released."
While they did not mention which manufacturers and chips will be affected by the update, it implies a collective approach to supporting China's domestic chipmakers.
FP8 (8-bit floating point) is a processing format which means the AI can run faster and use less memory than previous methods.
Additionally, DeepSeek added that V3.1 also features a hybrid inference structure, letting the model operate in both reasoning and non-reasoning modes.
A new “deep thinking” button on DeepSeek’s official app and web platform enables users to switch between these modes.
The company further announced changes to the pricing of its API services, which allow developers to integrate DeepSeek’s models into third-party apps and platforms. The updated pricing will take effect from September 6.
