Anthony Kwan | Getty Images News | Getty Images
Chinese artificial intelligence startup DeepSeek has hinted that China will soon have homegrown “next generation” chips to support its AI models, while announcing an update to one of its large language models.
In a comment under a post on its official WeChat account, DeepSeek said the “UE8M0 FP8” precision format of its newly released model V3.1 is tailored for the next-generation domestically built chips that will be launched soon.
FP8, or 8-bit floating point, is a data processing format that can boost the computational efficiency for training and inference of large deep learning models.
DeepSeek’s mention of China’s coming next-generation chips may signal plans to work more closely with China’s emerging AI chip ecosystem in the face of Washington’s advanced semiconductor export restrictions and Beijing’s push for chip self-sufficiency.
The comments come about two weeks after Beijing reportedly urged Chinese AI developers to use domestic alternatives to Nvidia’s graphics processing units used in AI training. While analysts say China’s domestic AI chipmakers lag behind Nvidia in technological advancement and scale, players like Huawei have been making progress.
In its Thursday post, DeepSeek did not disclose the chips it used to train the V3.1, or what local chips the UE8M0 FP8 might be compatible with.
DeepSeek shook up the tech world earlier this year after it released its R1 reasoning model, which demonstrated capabilities comparable to those of Western competitors like OpenAI, despite U.S. export controls restricting it from using Nvidia’s most advanced AI training chips.
Prior to that, in December, the company released its V3 model, which it said had been trained on about 2,000 of Nvidia’s less advanced chips.
Following DeepSeek’s model breakthroughs, the U.S. further tightened export restrictions in April, effectively banning Nvidia’s H20 chips, which had been specially designed to meet prior export restrictions on China.
Last month, officials from the Trump administration said they planned to allow Nvidia to resume shipping the chips to China. However, the H20s are now being met with scrutiny in China, with regulators reportedly mandating companies against buying the chips until a national security review is completed.
Chip analysts have told CNBC that companies like Huawei that have been seeking to build an alternative AI chip ecosystem in China could benefit from a lack of Nvidia’s H20s in the market.
DeepSeek said Thursday that its V3.1 came with “major changes,” including faster response times, and a hybrid reasoning architecture that allows the model to support both reasoning and non-reasoning modes. Reasoning models can execute more complicated tasks through a step-by-step logical thought process.
Starting Sept. 6, the company will also adjust the pricing for using the model’s API, which allows developers of other apps and web products to integrate DeepSeek on their platforms.