Posted in

DeepSeek-V3.1-Terminus

DeepSeek-V3.1-Terminus represents a significant update that maintains the model’s original capabilities while addressing key user-reported issues. The improvements focus on two main areas: language consistency and agent capabilities. Users will notice reduced instances of mixed Chinese-English text and fewer abnormal characters, making interactions smoother and more professional. Additionally, the performance of both Code Agent and Search Agent has been further optimized for enhanced functionality.

DeepSeek Logo

Badge

The benchmark results demonstrate impressive performance across multiple evaluation metrics:

Reasoning Mode without Tool Use:
– MMLU-Pro: 85.0
– GPQA-Diamond: 80.7
– Humanity’s Last Exam: 21.7
– LiveCodeBench: 74.9
– Codeforces: 2046
– Aider-Polyglot: 76.1

Agentic Tool Use:
– BrowseComp: 38.5
– BrowseComp-zh: 45.0
– SimpleQA: 96.8
– SWE Verified: 68.4
– SWE-bench Multilingual: 57.8
– Terminal-bench: 36.7

Chat Badge

Hugging Face Badge

The search agent has received significant updates with revised templates and tool-sets, detailed in the provided trajectory documentation. For developers looking to implement this model, the architecture remains consistent with DeepSeek-V3, making migration straightforward.

Discord Badge

WeChat Badge

Local deployment is supported through comprehensive documentation in the DeepSeek-V3 repository, with additional chat template information available in the DeepSeek-V3.1 repository. The inference folder contains updated demo code to help developers quickly understand and implement the model architecture.

Twitter Badge

It’s important to note a current technical consideration: the parameters of self_attn.o_proj in the current model checkpoint don’t conform to the UE8M0 FP8 scale data format. This known issue will be addressed in future releases.

License Badge

The model and repository operate under the MIT License, providing flexibility for commercial and research use. For academic purposes, the technical report citation is available for reference. The DeepSeek team maintains active support channels, encouraging users to reach out with questions through GitHub issues or direct email communication at service@deepseek.com.