DeepSeek hints latest model will be supported by China’s ‘next generation’ homegrown AI chips

Chinese AI startup DeepSeek has revealed that it is preparing to leverage domestically produced "next-generation" chips to enhance its AI models. This announcement coincides with the launch of an updated version of its large language model, V3.1. In a recent post on its official WeChat account, DeepSeek highlighted the significance of the "UE8M0 FP8" precision format, specifically designed for these upcoming homegrown chips. The FP8 format, which stands for 8-bit floating point, is a data processing method that significantly increases computational efficiency in the training and application of expansive deep learning models. The company's mention of new domestic chips suggests a strategic move to align with China's growing AI chip industry, especially in light of the U.S. semiconductor export restrictions. This comes after reports indicated that Beijing has encouraged local AI developers to turn to homegrown alternatives to Nvidia's GPUs for their AI training needs. While experts acknowledge that Chinese chipmakers are currently trailing Nvidia in both technology and production capacity, companies like Huawei are making strides in this arena. DeepSeek did not specify which chips were used for training the V3.1 model or which local chips might support the UE8M0 FP8 format. Earlier this year, DeepSeek made headlines with its R1 reasoning model, which exhibited performance levels comparable to leading Western counterparts such as OpenAI, despite being limited by U.S. export controls that restricted access to the most advanced Nvidia chips. In December, the company introduced its V3 model, trained on less advanced Nvidia chips. Following these advancements, the U.S. further tightened its export controls in April, effectively banning the H20 chips, which were tailored to comply with previous restrictions. Recent reports suggest that the Trump administration is reconsidering its stance on allowing Nvidia to resume shipments to China, although these H20 chips are now under scrutiny, with Chinese regulators requiring a national security review before any purchases can be made. Industry analysts have noted that firms like Huawei, which are working towards establishing a robust alternative AI chip ecosystem in China, may find opportunities due to the absence of Nvidia's H20 chips. DeepSeek also announced that its V3.1 model features significant enhancements, such as quicker response times and a hybrid reasoning architecture that accommodates both reasoning and non-reasoning functionalities. Starting September 6, the company will revise the pricing for its model's API, enabling developers to integrate DeepSeek's capabilities into their applications and web products.