Beijing (Reuters)-Launched it on Thursday an update of the V3 model which, according to the company, has a feature that can optimize it to chin chips in China, as well as faster processing speeds.
Focus on compatibility with household processors may indicate that Deepseek AI models are being positioned to work with the emerging ecosystem of China semiconductors as Beijing strives to replace US technology in the export restrictions of Washington.
Read more:
Take your business to the next level with the help of the country’s leading entrepreneurs!
Deepseek’s V3 model update follows two other recent renewals of its main models – an R1 update in May and an improvement of V3 in March.
To support home chips, DeepSeek said that the UE8M0 FP8 accuracy format of the Deepseek-V3.1 model is optimized for next-generation “next generation chips to be released soon”.
The company did not identify which models or manufacturers of specific chips would be compatible.
Continues after advertising
FP8, or 8 -bit floating point, is a data processing format that allows AI models to operate more efficiently using less memory and running commands faster than traditional methods.
DeepSek-V3.1 has a hybrid inference structure that allows the model to operate in reasoning and non-reasoning modes, the company said.
Users can switch between these modes using a “Deep Thinking” button on the company’s official application and on the web platform, which now perform the V3.1 version.
Continues after advertising
The company will also adjust the price for use of the model’s API, a platform that allows developers of other web applications and products to integrate their AI models from September 6.
