Particle.news
Download on the App Store
8 ARTICLES
·
yesterday

DeepSeek Releases V3.1, Tunes FP8 Format for Forthcoming Chinese AI Chips

Analysts view the FP8 change as a move toward tighter coordination with domestic hardware makers despite ongoing use of Nvidia systems for large‑scale training.

Image
Image
Image
Image

Overview

  • The V3.1 update adopts a UE8M0 FP8 precision format that DeepSeek says is tailored for next‑generation homegrown chips expected soon, though no vendors were named.
  • Built from a V3 checkpoint, the model consolidates reasoning and non‑reasoning into one system, with a new “deep thinking” toggle in DeepSeek’s app and web platform.
  • Context capacity doubles to 131,072 tokens and the company reports stronger tool and function calling, including big gains on autonomous browsing benchmarks.
  • Reporting indicates earlier attempts to train a next‑gen model on Huawei Ascend hardware struggled, leading DeepSeek to revert training to Nvidia H20 while evaluating Ascend for inference.
  • Model weights are available on Hugging Face and ModeScope, and DeepSeek plans to change API pricing for developers on Sept. 6 as Chinese chip and foundry shares rally on the news.