DeepSeek V3.1 Terminus — refined variant of V3.1 optimized for tool calling and structured generation tasks.
DeepSeek-V3.1-Terminus, released September 22, 2025, is a targeted refinement of V3.1 rather than a new architecture. It keeps the same 671B-total / 37B-active MoE backbone, 128K context, and hybrid-reasoning surface (deepseek-chat non-thinking vs deepseek-reasoner thinking modes). The release focuses on two engineering fixes reported by users: reduced Chinese-English language mixing and fewer spurious characters, plus measurable agentic improvements in Code Agent and Search Agent performance.
For TheRouter operators the practical value is stability: Terminus reduces the failure modes that previously required post-processing or prompt engineering. Agentic benchmarks show consistent gains — SWE-bench Verified 68.4 (+2.4), SWE-bench Multilingual 57.8 (+3.3), Terminal-bench 36.7 (+5.4), BrowseComp 38.5 (+8.5) — while reasoning-only benchmarks (GPQA-Diamond, HLE) move modestly. Pricing, context window, and API surface remain identical to V3.1. The model is MIT-licensed for code and permissively licensed for weights; weights are available on Hugging Face, and inference stacks (SGLang, vLLM, LMDeploy) already support it.
| Type | Rate |
|---|---|
| Input | $0.2916 / 1M tokens |
| Output | $1.08 / 1M tokens |
| Release date | 2025-09-22api-docs.deepseek.com ↗ | verified |
| Architecture | 671B-total / 37B-active MoE (unchanged from V3.1); same tokenizer and chat template; targeted post-training for language consistency and agent reliabilityapi-docs.deepseek.com ↗ | verified |
| Context window | 128K tokensapi-docs.deepseek.com ↗ | verified |
| Hybrid modes | deepseek-chat → non-thinking mode; deepseek-reasoner → thinking mode (toggle via API or UI DeepThink button). Tool calls in reasoner mode are automatically rerouted to chat mode.venturebeat.com ↗ | verified |
| License — code | MIThuggingface.co ↗ | verified |
| License — weights | DeepSeek Model License (commercial use permitted)huggingface.co ↗ | verified |
| Supported inference backends | SGLang, vLLM, LMDeploy, TensorRT-LLM, LightLLM, DeepSeek-Infer; FP8 + BF16; NVIDIA + AMD GPUs; Huawei Ascend via MindIEhuggingface.co ↗ | verified |
| Successor | DeepSeek-V3.2 (Dec 2025) → V3.2-Speciale (temporary) → V4-Flash / V4-Pro (Apr 2026) | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-bench Verified Pass@1; +2.4 over V3.1 (66.0) | 68.4% | venturebeat.com ↗ | |
SWE-bench Multilingual Pass@1; +3.3 over V3.1 (54.5) | 57.8% | venturebeat.com ↗ | |
Terminal-bench +5.4 over V3.1 (31.3) | 36.7% | venturebeat.com ↗ | |
BrowseComp +8.5 over V3.1 (30.0) | 38.5% | venturebeat.com ↗ | |
SimpleQA +3.4 over V3.1 (93.4) | 96.8% | venturebeat.com ↗ | |
MMLU-Pro +0.2 over V3.1 (84.8) | 85.0% | sourceforge.net ↗ | |
GPQA-Diamond +0.6 over V3.1 (80.1) | 80.7% | sourceforge.net ↗ | |
Humanity's Last Exam +5.8 over V3.1 (15.9) | 21.7% | venturebeat.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "deepseek/deepseek-v3.1-terminus",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"deepseek/deepseek-v3.1-terminus","messages":[{"role":"user","content":"Hello"}]}'Minor update to V3.1 that reduces CN/EN mixing and spurious characters while delivering measurable gains on SWE-bench, BrowseComp, and Terminal-bench. Same pricing and context window; aliases deepseek-chat and deepseek-reasoner now point to Terminus.
It is a refinement checkpoint of the same V3.1 architecture. DeepSeek performed additional post-training focused on language consistency and agent reliability. Weights are published separately on Hugging Face as DeepSeek-V3.1-Terminus.
Yes for any production workload that previously experienced language mixing or unstable tool calling. The upgrade is drop-in (same model ID surface via TheRouter) and carries no additional cost. New projects should evaluate V4-Flash / V4-Pro first.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| Architecture | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| Context window | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| Hybrid modes | venturebeat.com ↗ | 2026-05-25 | verified |
| License — code | huggingface.co ↗ | 2026-05-25 | verified |
| License — weights | huggingface.co ↗ | 2026-05-25 | verified |
| Supported inference backends | huggingface.co ↗ | 2026-05-25 | verified |
| Successor | — | — | verified |
| SWE-bench Verified | venturebeat.com ↗ | 2026-05-25 | verified |
| SWE-bench Multilingual | venturebeat.com ↗ | 2026-05-25 | verified |
| Terminal-bench | venturebeat.com ↗ | 2026-05-25 | verified |
| BrowseComp | venturebeat.com ↗ | 2026-05-25 | verified |
| SimpleQA | venturebeat.com ↗ | 2026-05-25 | verified |
| MMLU-Pro | sourceforge.net ↗ | 2026-05-25 | verified |
| GPQA-Diamond | sourceforge.net ↗ | 2026-05-25 | verified |
| Humanity's Last Exam | venturebeat.com ↗ | 2026-05-25 | verified |
| DeepSeek releases V3.1-Terminus — language consistency & agent reliability fixes | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| Is V3.1-Terminus a new model or just a patch? | api-docs.deepseek.com ↗ | 2026-05-25 | to verify |
| Should I migrate from V3.1 to V3.1-Terminus? | api-docs.deepseek.com ↗ | 2026-05-25 | to verify |