AsianFin -- Beijing-based AI firm DeepSeek on Wednesday announced the release of DeepSeek-V3.1, highlighting major technical upgrades in model precision and architecture.

In its official statement, the company said the new version adopts UE8M0 FP8 Scale parameter precision, a move designed to optimize performance and efficiency. DeepSeek also introduced significant changes to its tokenizer and chat template, setting V3.1 apart from its predecessor, V3.

Notably, in a pinned comment on its official WeChat account, DeepSeek revealed that the UE8M0 FP8 format was specifically designed for the next generation of domestically developed chips, signaling closer alignment between large-scale AI models and China’s homegrown semiconductor ecosystem.