Direct answer
DeepSeek Terminus usually refers to DeepSeek-V3.1-Terminus, DeepSeek’s September 2025 update to V3.1. DeepSeek said it targeted fewer mixed Chinese-English outputs and abnormal characters, plus improvements to Code Agent and Search Agent performance. Those are the publisher’s stated aims—not a guarantee that it will outperform other models on your tasks. (deepseek.com)
For a new production integration with DeepSeek’s own API, don’t assume Terminus is a currently selectable model. DeepSeek’s special Terminus comparison endpoint was explicitly temporary, with access ending October 15, 2025. Current API documentation lists newer model options, including deepseek-flash and deepseek-v4-pro; it does not establish Terminus as a current selectable API model. Terminus weights remain published on Hugging Face, but self-hosting or a third-party provider is a different access route with its own requirements and terms. (api-docs.deepseek.com)
Practical choice: consider Terminus if you specifically need to evaluate or self-host that V3.1 checkpoint and can verify your serving route. For a new DeepSeek-hosted API deployment, evaluate the model IDs currently offered in DeepSeek’s API documentation instead; do not copy the retired Terminus endpoint into production.
What Terminus changed—and what the evidence says
DeepSeek described Terminus as an update to V3.1, intended to improve language consistency and Code Agent and Search Agent performance. Its official model card publishes a comparison against V3.1. For example, it reports SWE-bench Verified at 68.4 versus 66.0 and Terminal-Bench at 36.7 versus 31.3 in its agentic-tool-use results. But not every score improved: the card reports Codeforces at 2,046 versus 2,091 and BrowseComp-zh at 45.0 versus 49.2. These are publisher-provided benchmark results, not a substitute for testing your workload or proof of a general performance advantage. (huggingface.co)
| Your priority | What the available evidence supports | Decision implication |
|---|---|---|
| Coding or tool-using workflows | DeepSeek reports improvements on some agent-use benchmarks, including SWE-bench Verified and Terminal-Bench. | Put Terminus on a test shortlist if you can run the weights or reach them through a verified provider. Test your actual tools and tasks. |
| Consistent English/Chinese output | DeepSeek says the update reduces mixed-language output and abnormal characters. | Treat this as a release claim to validate on your prompts, especially if mixed-language responses are a known problem. |
| Broadly better general reasoning | The published comparison is mixed across measures; for instance, Codeforces is lower for Terminus than V3.1 in the model card. | Don’t choose it based on the Terminus name or one favorable score. Compare on your task set. |
| A directly selectable DeepSeek API model today | The dedicated Terminus comparison route had a stated end date in 2025; current API docs describe newer model choices. | Verify the exact current model ID and route before designing around Terminus. (api-docs.deepseek.com) |
Production API: check the access route first
DeepSeek announced Terminus as available through its API in 2025, and its later V3.2-Exp announcement offered a temporary Terminus comparison endpoint ending October 15, 2025. That historical availability is not evidence that the same endpoint is a supported production route now. The old sample uses a URL containing v3.1_terminus_expires_on_20251015; its own date is a clear reason not to build a new integration around it. (deepseek.com)
Before selecting Terminus, verify all of the following for the specific route you intend to use:
- Model listing and identifier: Is Terminus explicitly offered now by DeepSeek or by your chosen host? Record the exact model ID—not just “DeepSeek” or “Terminus.”
- Route and lifecycle: Confirm the documented base URL, whether the endpoint is production-supported, and any retirement or expiry date.
- Feature compatibility: Check the request format, tool/function-calling support, context and output limits, and streaming behavior that your application needs. Don’t transfer V3.1 or other model specifications to Terminus without documentation that they apply.
- Provider-specific terms: If using a host rather than DeepSeek directly, check that host’s availability, data handling, limits, service commitments, and billing. A published Hugging Face checkpoint confirms the weights are available there; it does not by itself confirm a production API or equivalent service terms. (
huggingface.co) - Cost: Compare current input and output pricing for the exact model and route you will call. The historical comparison announcement said Terminus calls were priced like V3.2-Exp during that temporary test; it is not a current price quote. (
api-docs.deepseek.com)
DeepSeek’s current API quick-start documentation shows its present API configuration and model examples. Use the current documentation for a new integration rather than the old Terminus sample. (api-docs.deepseek.com)
A useful workload evaluation
If a provider confirms that it still serves Terminus, compare it with the model you would otherwise deploy using the same fixed test set. Include representative coding or tool-use tasks only if those matter to your application; include English, Chinese, or mixed-language prompts if language consistency is a concern. Define success before testing—for example, task completion, correct tool selection, valid structured output, and unacceptable language-switch or character errors.
Run the same prompts and tool definitions through both models, then compare success rate, latency, and token use under your expected settings. Repeat variable tasks and inspect failures, not just average scores. This is a proposed evaluation method, not a claim that such a production test has been run. It helps determine whether Terminus’s reported strengths matter for your workload enough to justify the serving, compatibility, and maintenance trade-offs.
Bottom line
DeepSeek-V3.1-Terminus is a real, officially announced V3.1 update with publisher-reported gains in language consistency and some agent benchmarks. However, its special comparison API was temporary and its documented expiry has passed. For a new DeepSeek API deployment, choose from the model IDs currently listed in DeepSeek’s API documentation unless DeepSeek confirms a supported Terminus route for your account. If you can access the weights through self-hosting or a third-party provider, treat Terminus as a workload-specific candidate and validate the exact checkpoint, serving route, terms, and results before production use. (deepseek.com)
Status note: this article reflects the documentation available on November 9, 2026. Model listings and API terms can change; verify the linked documentation when making an implementation decision.
Sources
DeepSeek: DeepSeek-V3.1 is now DeepSeek-V3.1-Terminus
DeepSeek-V3.1-Terminus model card
DeepSeek API: Terminus chat sample