The Deepseek logo is seen in this illustration taken on January 29, 2025. REUTERS/Dado Ruvic/Illustration Purchase LicensingRights
BEIJING, Aug 21 (Reuters) - Chinese artificial intelligence startup DeepSeek released on Thursday an upgrade to its flagship V3 model that the company says has a feature that can optimize it for Chinese-made chips, along with faster processing speeds.
The focus on domestic chip compatibility may signal that DeepSeek's AI models are being positioned to work with China's emerging semiconductor ecosystem as Beijing pushes to replace U.S. technology in the face of Washington's export restrictions.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign uphere.
00:16AI Weekly: grouchy Grok and the wizard of bots
DeepSeek shook the technology world this year when it released AI models that compete with Western ones like OpenAI's ChatGPT while offering lower operational costs.
The upgrade to DeepSeek's V3 model follows two other recent updates to its core models - an R1 model update in May and an earlier V3 enhancement in March.
For domestic chip support, DeepSeek said in a WeChat post its DeepSeek-V3.1 model's UE8M0 FP8 precision format is optimised for "soon-to-be-released next-generation domestic chips".
Advertisement · Scroll to continue
The company did not identify which specific chip models or manufacturers would be supported.
FP8, or 8-bit floating point, is a data processing format that allows AI models to operate more efficiently, using less memory while running faster than traditional methods.
The DeepSeek-V3.1 features a hybrid inference structure that enables the model to operate in both reasoning and non-reasoning modes, the company said in a WeChat post on Thursday.
Advertisement · Scroll to continue
Users can toggle between these modes using a "deep thinking" button on the company's official app and web platform, both of which now run the V3.1 version.
The company will also adjust the costs for using the model's API, a platform that allows developers of other apps and web products to integrate its AI models, starting September 6, the statement showed.
