Chinese DeepSeek Releases Updated Model Tailored for Domestic Chips

Chinese artificial intelligence company DeepSeek announced a new version of its flagship model V3 on Thursday, emphasizing that the new model, in addition to increased data processing speed, also features an optimization capability for chips manufactured in China.

The focus on compatibility with Chinese chips could signal DeepSeek’s efforts to integrate its artificial intelligence models into the advanced chip system currently being developed in China, as official Beijing seeks to replace American technology due to export restrictions imposed by Washington.

Earlier this year DeepSeek shook the tech sector with its AI models that offer lower operational costs, along with reduced consumption of electricity, water, and chips, competing with Western models such as ChatGPT from the American OpenAI.

The new version is the third in a series, following the earlier update of model V3 in March and the update of model R1 in May.

Regarding the adaptation to locally produced chips, DeepSeek announced on the social network WeChat that the data processing format UE8M0 FP8, on which the DeepSeek-V3.1 model operates, is optimized for “domestic next-generation chips that will soon be released to the market.”

The company did not specify which particular chip models or manufacturers will be supported.

The term FP8 refers to a data processing format that allows artificial intelligence models to operate more efficiently, as it uses less memory and works faster than traditional methods.

The DeepSeek-V3.1 version features a hybrid reasoning structure, which allows the model to operate in two modes, with and without reasoning, the company announced on Thursday on the WeChat social network.

Users can toggle the two modes on and off using the “deep thinking” button in the official app and on the company’s online platform, where version V3.1 is now in use.

Additionally, the company will adjust the pricing for API usage starting September 6, a platform that allows developers to integrate DeepSeek’s artificial intelligence models into their products, according to a statement.