Deepseek hints at the latest model compatible with the next -generation home chips in China

The Chinese launch of the Deepseek artificial intelligence hinted that China will soon be able to “next -generation” homegrown chips to support its AI models, announcing an update to one of its large language models.

In the comment according to the message about it Official WeChat accountDeepseek said that “UE8m0 FP8” format its recently released v3.1 model is designed for the built -in -built -in neighbors to be launched soon.

FP8, or 8-bit floating, this Data processing format This can increase computing efficiency for learning and conclusions of great deep learning models.

DEEPEEK’s mention of the next generation’s coming chips may signal the plans to work more closely with the new chips in China in the face of export restrictions in a semiconductor semiconductor and the press of Beijing to the self -sufficiency of the chip.

The comments come about two weeks after Beijing has been reported to use Chinese developers to use internal alternatives to Nvidia graphic units used in AI learning. While analysts say Achieving progress.

In his message on Thursday, Deepseek did not reveal the chips he used to train v3.1, or about the local UE8M0 FP8 chips may be compatible.

Deepseek shook the world of technology earlier this year after releasing its R1 reasoning, which demonstrated opportunities comparable to Western competitors, such as Openai, despite the fact that US export control restricts the use of Nvidia’s most advanced training chips.

Prior to that, in December, the company released its model V3, which, she said, was preparing about 2000 with less advanced Nvidia chips.

Following the breakdown of the Deepseek model in April, in April, the export restrictions were further strengthened, actually banning H20 Nvidia chips, which were specifically designed to implement previous export restrictions on China.

Last month, Trump administration officials said Nvidia plan to restore chips to China. However, the H20 is now meeting with close attention in China, with the regulators, reportedly reported Mounting companies from purchase Chips until the National Security Review is completed.

Analysts chips said CNBC that companies like companies Huawei The sought to build an alternative Ecosystem of AI chips in China can take advantage of the H20s Nuji lack.

Deepseek said on Thursday that his v3.1 came with “serious changes”, including a faster response time, and hybrid reasoning architecture, which allows the model to support both reasoning and non -calling modes. Models of reasoning can perform more complex tasks through a step -by -step logical process of thinking.

Starting on September 6, the company also set up prices for the use of API model that allows developers of other applications and web products to integrate Deepseek on its platforms.