DeepSeek-V3.1 Release Notes

DeepSeek-V3.1 is a model update designed to transition toward the "agent era" by introducing hybrid inference modes and enhanced tool-use capabilities. The release focuses on reducing thinking latency, improving multi-step reasoning, and extending context windows for both thinking and non-thinking modes.

Hybrid Inference: Think and Non-Think Modes

DeepSeek-V3.1 implements a hybrid inference system that allows users to toggle between two distinct modes of operation within a single model architecture:

  • Thinking Mode (deepseek-reasoner): This mode is optimized for complex reasoning. DeepSeek-V3.1-Think is reported to reach answers faster than the previous DeepSeek-R1-0528 model.
  • Non-Thinking Mode (deepseek-chat): This mode provides standard chat interactions without the extended reasoning process.

Both modes support a 128K context window.

Agentic Capabilities and Tool Use

DeepSeek-V3.1 features post-training enhancements specifically aimed at boosting tool use and multi-step agent tasks. These upgrades result in:

  • Improved Benchmarks: The model shows better results on SWE and Terminal-Bench.
  • Complex Search: There are stronger multi-step reasoning capabilities for complex search tasks.
  • Efficiency: The model demonstrates significant gains in thinking efficiency.

To support these agentic workflows, the Beta API now supports Strict Function Calling.

Technical Model Updates

DeepSeek-V3.1 is built upon the V3 architecture with the following technical modifications:

  • Continued Pretraining: The V3.1 Base model underwent continued pretraining with 840B tokens to extend long context capabilities.
  • Tokenizer and Template: The tokenizer configuration and chat templates have been updated.
  • Open Source Availability: Both the V3.1 Base and the standard V3.1 weights are available via Hugging Face.

API and Integration Details

DeepSeek has updated its API to align with the new model modes and updated its integration options:

  • API Endpoints: deepseek-chat is used for non-thinking mode and deepseek-reasoner for thinking mode.
  • Compatibility: The API now supports the Anthropic API format.
  • Infrastructure: DeepSeek has increased API resources to ensure a smoother user experience.

Pricing and Timeline

New pricing structures and the end of off-peak discounts will take effect on September 5, 2025, at 16:00 UTC. Until that date, the current API pricing remains in effect.

Sources

Related

  • Dispatch
    DeepSeek-V3.2 and DeepSeek-V3.2-Speciale Release

    DeepSeek has released DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, reasoning-first models designed for agents that introduce integrated thinking in tool-use and gold-medal performance in competitive programming and mathematics.

  • Dispatch
    DeepSeek-V3.1-Terminus Release Notes

    DeepSeek has released DeepSeek-V3.1-Terminus, an update to V3.1 that improves language consistency and enhances the performance of Code and Search Agents.

  • Dispatch
    DeepSeek-V4-Pro GA Release

    DeepSeek has released DeepSeek-V4-Pro, featuring enhanced agent capabilities, flexible reasoning effort settings, and native OpenAI Responses API support.

  • Dispatch
    DeepSeek-V4 Preview Release

    DeepSeek has released the open-source DeepSeek-V4 preview, introducing two model variants—Pro and Flash—featuring a 1M token context window and state-of-the-art agentic coding capabilities.

  • Dispatch
    DeepSeek-V4 Release Notes: Efficient 1M-Token Context for AI Agents

    DeepSeek-V4 introduces a 1M-token context window powered by a hybrid CSA/HCA attention mechanism, specifically optimized for long-running agentic workloads and tool-use trajectories.