DeepSeek-V3.1
Hybrid thinking and non-thinking language model with 671 billion parameters.
DeepSeek-V3.1 is a large language model released by DeepSeek AI that supports both reasoning and direct response modes. It features 671 billion parameters, uses a mixture of experts architecture, and offers open weights under an MIT licence for research and commercial use.
- 671 billion total parameters with 37 billion active per token
- Supports both thinking and non-thinking response modes
- Context window extends to 128,000 tokens
- Licensed under MIT for open use
- Uses UE8M0 FP8 format for weights and activations
What it is
DeepSeek-V3.1 is a large language model developed by DeepSeek AI. It contains 671 billion parameters in total, with 37 billion active during any single inference step. The model supports two distinct operational modes: a thinking mode for complex reasoning tasks and a non-thinking mode for faster, direct responses. It handles context windows up to 128,000 tokens and includes improved capabilities for using external tools and agents. The model is built upon the original V3 base checkpoint and uses a hybrid architecture to balance speed and accuracy.
How it came about
The model represents an upgrade to the previous DeepSeek-V3 version. Developers extended the training process significantly compared to earlier iterations. The phase for extending context to 32,000 tokens increased tenfold to 630 billion tokens. The subsequent phase for 128,000 token context grew by 3.3 times to 209 billion tokens. This extensive post-training aimed to enhance long-document handling and overall performance stability. The release follows the methodology established in the original technical report for the V3 series.
Inside the model
DeepSeek-V3.1 employs a mixture of experts architecture, which allows it to activate only a subset of its parameters for each token. This design improves efficiency while maintaining high capacity. The model uses UE8M0 FP8 scale data formats for both weights and activations to ensure compatibility with microscaling standards. A key architectural feature is the hybrid thinking mode, controlled via chat templates. Users can switch between reasoning-heavy outputs and standard responses by adjusting the prompt structure. Tool calling is integrated into the non-thinking mode, supporting various agent frameworks.
What it can do
In benchmark tests, the thinking mode achieves scores comparable to the DeepSeek-R1-0528 model. On MMLU-Redux, it scores 93.7, and on GPQA-Diamond, it reaches 80.1. The non-thinking mode scores 91.8 on MMLU-Redux and 74.9 on GPQA-Diamond. For coding tasks, the thinking mode attains a rating of 2091 on Codeforces-Div1 and 74.8 on LiveCodeBench. Agent-based evaluations show strong performance, with 66.0 on SWE Verified and 54.5 on SWE-bench Multilingual. The model also demonstrates improved search agent capabilities through multi-turn tool calling.
How to use it
The model weights are available for download under the MIT Licence. Users can access the base and chat versions through HuggingFace and ModelScope platforms. Running the model locally requires specific precision settings, particularly for gate parameters. The developers provide detailed chat templates for implementing both thinking and non-thinking modes. Tool call formats are documented for integration with code and search agents. Compatibility with FP8 formats is emphasized for efficient deployment on supported hardware.
Questions people ask about DeepSeek-V3.1
What is the difference between thinking and non-thinking modes?
Thinking mode is designed for complex reasoning tasks and produces higher quality answers on difficult benchmarks. Non-thinking mode provides faster responses and supports tool calling. Users switch between these modes by changing the chat template prefix.
How many parameters does DeepSeek-V3.1 have?
The model has a total of 671 billion parameters. However, it uses a mixture of experts architecture, so only 37 billion parameters are activated for each token processed. This design balances computational efficiency with model capacity.
Can I use DeepSeek-V3.1 for commercial purposes?
Yes, the model weights are released under the MIT License. This permissive licence allows for both research and commercial usage. You can download the weights from HuggingFace or ModelScope and integrate them into your applications.
What is the maximum context length supported?
DeepSeek-V3.1 supports a context length of 128,000 tokens. This allows the model to process and reason over very long documents. The training process included specific phases to extend context capabilities from 32K to 128K tokens.
Good reading and sources
There is no English Wikipedia article on DeepSeek-V3.1 yet, so the facts on this page come from the sources listed above: huggingface.co/deepseek-ai/DeepSeek-V3.1. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.