DeepSeek-V3.1-Terminus

// Overview

DeepSeek-V3.1-Terminus is an incremental but practical update to DeepSeek's V3.1 line, released by the Chinese AI lab DeepSeek on September 22, 2025 under an MIT license with open weights on Hugging Face. Rather than a new architecture, Terminus is a refinement: it keeps V3.1's 685B-parameter Mixture-of-Experts design (about 37B parameters active per token) and its hybrid thinking/non-thinking modes, while fixing real-world rough edges users had reported.

The two headline fixes are language consistency and agent quality. DeepSeek-V3.1-Terminus produces far fewer Chinese/English mix-ups and largely eliminates the random stray-character glitches that occasionally appeared in V3.1, giving more coherent multilingual output. On top of that, its built-in Code Agent and Search Agent were tuned to be more reliable inside tool-using and agentic workflows.

The model is text-only and exposes a 128K-class context window (163,840 tokens) with up to 32,768 output tokens. Because it ships as open weights under MIT, DeepSeek-V3.1-Terminus can be self-hosted or accessed through DeepSeek's own API and a range of third-party providers, making it a low-cost option for agentic coding, retrieval and search tasks.

// Benchmarks

1. MMLU-Pro85%
2. GPQA-Diamond80.7%
3. Humanity's Last Exam21.7%
4. LiveCodeBench74.9%
5. Aider-Polyglot76.1%
6. SWE-bench Verified68.4%
7. SWE-bench Multilingual57.8%
8. Terminal-bench36.7%
9. BrowseComp38.5%
10. BrowseComp-zh45%
11. SimpleQA96.8%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Pricing shown is the listed rate for DeepSeek-V3.1-Terminus on OpenRouter; the model also ships as open weights for self-hosting. Third-party and official DeepSeek API rates vary.

Pricing source ↗

// Strengths

• Open weights under a permissive MIT license — free to self-host, fine-tune, and deploy commercially
• Hybrid model: one set of weights serves both a fast non-thinking chat mode and a deeper reasoning mode
• Strong agentic coding and tool-use scores (SWE-bench Verified 68.4, Terminal-bench 36.7)
• Improved search/browse agent (BrowseComp 38.5) and very high factual recall (SimpleQA 96.8)
• Cleaner multilingual output — fixes V3.1's language mixing and stray-character issues
• Low API pricing relative to comparable frontier models, plus self-hosting option

// Best for

• Agentic coding assistants and autonomous software-engineering loops (SWE-bench-style tasks)
• Terminal and CLI agents that execute multi-step commands
• Web search and browsing agents that gather and synthesize information
• Multilingual chat and content generation with consistent language output
• Long-context document analysis up to 128K tokens
• Cost-sensitive or privacy-sensitive deployments via self-hosted open weights

// How to access

// DeepSeek V3 — every version

The full lineage of the DeepSeek V3 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.