DeepSeek‑V3.1‑Terminus Fixes the ‘Extreme’ Bug and Outperforms Gemini 2.5 Pro

DeepSeek‑V3.1‑Terminus released

DeepSeek announced the new model DeepSeek‑V3.1‑Terminus, which addresses the previously reported “extreme” character bug and improves language consistency, especially reducing mixed Chinese‑English output and occasional abnormal characters.

All official channels (App, web, mini‑program, and API) have been updated to the new model.

Download links:

Hugging Face: https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Terminus

ModelScope: https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.1-Terminus

The update also enhances Agent capabilities, optimizing both Code Agent and Search Agent performance.

Benchmark results show significant gains over the previous version and over Gemini 2.5 Pro on several tasks, including a 36.48 % increase on Humanity’s Last Exam, improvements on MMLU‑Pro, GPQA‑Diamond, LiveCodeBench, SimpleQA, and various SWE‑bench scores.

Agent‑related scores also rose: BrowseComp 30.0 → 38.5, SimpleQA 93.4 → 96.8, SWE‑bench Verified 66.0 → 68.4, SWE‑bench Multilingual 54.5 → 57.8, Terminal‑bench 31.3 → 36.7.

The “extreme” bug that inserted the character “极” into outputs (e.g., turning time.Second into time.Se极) appears to be resolved; the new model provides five correct timer implementations without the anomaly.

Some benchmarks (Codeforces, Aider‑Polyglot, BrowseComp‑zh) showed slight declines, but overall performance is markedly better.

Speculation about the next model, DeepSeek‑V4/R2, is ongoing, with the community eagerly awaiting further announcements.

References:

https://x.com/deepseek_ai/status/1970117808035074215

https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Terminus

https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.1-Terminus