The 685-billion-parameter model introduces what DeepSeek calls a “hybrid inference structure,” allowing users to toggle between rapid “non-thinking” responses for basic queries and slower “thinking” mode for complex reasoning tasks through a “DeepThink” button on the platform. According to the company’s WeChat announcement, this represents “our first step toward the agent era”.

DeepSeek-V3.1 delivers significant performance improvements over its predecessors, scoring 71.6% on the prestigious Aider coding benchmark and achieving comparable results to proprietary models like Anthropic’s Claude Opus while costing substantially less. The model processes tasks at approximately $1.01 each, compared to nearly $70 for equivalent workloads from competing systems.

The release coincides with DeepSeek’s preparation for next-generation Chinese-made AI chips, utilizing FP8 (8-bit floating point) processing formats optimized for domestic hardware. This technical specification suggests the company is positioning itself for China’s emerging semiconductor ecosystem as Beijing pushes to reduce dependence on Western chip manufacturers.

The timing appears deliberate, coming as the Trump administration recently permitted Nvidia to resume limited AI chip sales to China through its H20 model, though Beijing’s response has been notably tepid. Chinese officials have advised domestic firms against relying on the American chips, reflecting broader ambitions for technological self-sufficiency.