DeepSeek V3.1 Debuts Hybrid Reasoning AI Model

DeepSeek has released version V3.1 of its large language model, introducing a hybrid reasoning architecture that allows the system to switch between “thinking” and “non-thinking” modes. The update is designed to improve reasoning performance, multi-step task execution, and tool use. While maintaining practical inference speeds for production environments. 

The new DeepSeek-V3.1-Think mode delivers faster reasoning compared to the earlier DeepSeek-R1-0528 model, with similar response quality. The model leverages a two-phase context extension strategy, expanding input length first to 32,000 tokens with 630 billion training tokens, and later to 128,000 tokens using an additional 209 billion tokens. This enables significantly longer sequence handling than prior releases. 

Technically, DeepSeek V3.1 uses FP8 UE8M0 precision for weights and activations, improving efficiency and maintaining compatibility with microscaling techniques. The full model contains 671 billion parameters, with about 37 billion activated per token. 

Early benchmarks place V3.1 near the top of open-source AI models. On Aider benchmark, it scored 71.6%, outperforming Anthropic’s Claude 4 and approaching OpenAI’s GPT-4, while costing roughly $1 in compute per full run—far below the cost of rival systems. 

Community feedback has been positive, with developers citing its cost-performance advantage and strong coding capabilities. One user called it “insane” for code reviews. Others praised the hybrid inference design for balancing deep reasoning with quick responses. 

The model is accessible via Hugging Face, OpenRouter, and Replicate, alongside detailed API documentation and release notes. By integrating structured post-training improvements with hybrid inference, DeepSeek aims to position V3.1 as a cost-effective, high-performance alternative to premium closed-source models. 

For enterprises and developers, the release underscores the growing trend toward scalable, reasoning-capable AI models that merge advanced context handling with efficiency for real-world applications. 

 

Source: 

https://www.infoq.com/news/2025/09/deepseek-v31-hybrid/