China's DeepSeek Unveils V3.1, a 685B Model for Domestic Chips

China’s DeepSeek has quietly released its largest and most ambitious model to date, V3.1, a 685-billion-parameter system optimized to run on domestically produced chips. The update, unveiled via a low-profile post on the company’s platform and on an open-source hub, combines a hybrid reasoning architecture with a dramatically expanded context window and a lower-cost profile designed to undercut Western frontier models. Early benchmark results place V3.1 in close competition with leading proprietary systems, while its open-source licensing aims to accelerate adoption around the world, particularly in regions seeking alternatives to closed, high-cost AI stacks.

Technically, V3.1 represents a substantial evolution for the Hangzhou-based startup. It retains the core V3 lineage but enhances performance through a mixture-of-experts design that activates only a portion of the model for each query, reducing compute costs. The model supports multiple numeric precisions, including BF16, FP8, and F32, and pushes context length up to 128,000 tokens, enabling analysis of long documents and large coding tasks. A key feature is its hybrid architecture, which fuses conversational capabilities...