Anthony Kwan | Getty Images News | Getty images
The Chinese artificial intelligence startup Deepseek has suggested that China will soon have “next generation” harvest “chips to support its AI models, while announcing an update of one of its large language models.
In a comment under a publication about his official Wechat account, Depseek said that the “EU8M0 FP8” precision format of its V3.1 model recently launched to the next generation of chips built nationwide that will be launched soon.
FP8, or 8 -bit floating point, is a data processing format that can increase computing efficiency for training and inference of large deep learning models.
Deepseek’s mention of China’s next next generation chips can indicate plans to work more closely with the Chinese AI chips ecosystem against the advanced export restrictions of Washington semiconductors and Beijing’s impulse due to the self -sufficiency of chip.
The comments occur approximately two weeks after Beijing urged Chinese developers to use national alternatives to NVIDIA graphics processing units used in AI training. While analysts say that Chinese chips manufacturers of China are left behind Nvidia in the advancement and technological scale, players like Huawei have been progressing.
In his Thursday publication, Depseek did not reveal the chips he used to train V3.1, or with what local chips the EU8M0 FP8 could be compatible.
Deepseek shook the world of technology earlier this year after it launched its R1 reasoning model, which demonstrated capabilities comparable to those of Western competitors such as OpenAi, despite the export controls of the United States that restrict the use of training chips of the most advanced NVIDIA.
Before that, in December, the company launched its V3 model, which according to him had been trained in approximately 2,000 of the less advanced Nvidia chips.
After the Deepseek model advances, the United States further hardened export restrictions in April, effectively prohibiting NVIDIA H20 chips, which had been specially designed to meet previous export restrictions in China.
Last month, Trump administration officials said they planned to allow Nvidia to resume the sending of the chips to China. However, the H20 are now being fulfilled with scrutiny in China, and according to reports, regulators demand companies that do not buy chips until a national security review is completed.
Chips analysts have told CNBC that companies such as Huawei that have been looking to build an alternative the chips ecosystem in China could benefit from the lack of NVIDIA H20 in the market.
Deepseek said Thursday that its V3.1 arrived with “important changes”, which include faster response times and a hybrid reasoning architecture that allows the model to support both modes of reasoning and not condition. Reasoning models can execute more complicated tasks through a logical thinking process step by step.
As of September 6, the company will also adjust the price of using the model’s API, which allows developers from other applications and web products to integrate Deepseek into their platforms.
