DeepSeek Unveils New AI Model Designed for Local Chinese Chips

The Chinese startup "DeepSeek" DeepSeek has revealed its development of local "next-generation" chips to support AI models, as part of the country's efforts to enhance reliance on local technology in the face of restrictions on importing advanced chips.

The company explained in a post on its official account on the WeChat platform that its new language model V3.1 relies on the precision format "UE8M0 FP8", specifically designed for the upcoming local next-generation chips set to be launched soon. The FP8, or 8-bit floating point, is a data processing format that enhances computational efficiency for training and inference of large deep learning models.

This announcement comes as part of China's efforts to enhance technological self-sufficiency, especially after Beijing urged local AI developers to use local alternatives to NVIDIA graphics processing units, which are critical for training advanced models.

DeepSeek commented in its post saying: "The precision format UE8M0 FP8 is specifically designed for locally manufactured next-generation chips that will be launched soon." This trend is believed to indicate closer collaboration with the emerging AI chip ecosystem in China, although the chips used for training version 3.1 or the compatible local chips with UE8M0 FP8 have not been disclosed.

It is worth noting that "DeepSeek" made waves in the tech sector earlier this year with the launch of its model R1, which demonstrated performance competing with models from Western companies such as OpenAI, despite U.S. restrictions preventing the use of the latest NVIDIA chips.

Chinese companies like Huawei, which are seeking to build an alternative AI ecosystem, are expected to benefit from the shortage of H20 processors from NVIDIA in the market, according to analysts.

DeepSeek also announced that its release V3.1 includes significant improvements, including "faster response times, and a hybrid inference architecture that allows the model to support both inference and non-inference modes," enabling it to perform more complex tasks through a gradual logical reasoning process.

Starting from September 6, the company will also adjust the pricing for using the model's API, allowing application developers and web products to integrate "DeepSeek" into their platforms more easily.