DeepSeek Launches v3.1 Model with Hybrid Reasoning and Competitive Pricing

DeepSeek has unveiled its v3.1 model, featuring a hybrid reasoning architecture that combines thinking and non-thinking modes. The update enhances agent capabilities, tool use, and task execution, positioning it as a formidable player in the AI space. Users can toggle between modes via a "Deep Thinking" button, mirroring functionality seen in Anthropic's Opus and Sonnet models.

Benchmark performance shows improvements in SWE and Terminal-Bench metrics, with the model scoring 60 on Artificial Analysis' intelligence index—just ahead of its predecessor, R1. Despite maintaining the same underlying architecture (671B total parameters, 37B active), v3.1 demonstrates higher efficiency and reduced token usage in reasoning mode. However, it trails Alibaba's latest offering and OpenAI's GPT-OSS in raw performance, while the absence of function calling in reasoning mode remains a limitation for agentic workflows.

The model debuted on Hugging Face earlier this week before rolling out to DeepSeek's official platforms. Pricing adjustments suggest a strategic push for broader adoption amid intensifying competition in the generative AI sector.