Short answer: DeepSeek described V3.1-Terminus as an update to V3.1 focused on language consistency and Code Agent and Search Agent performance. DeepSeek’s official announcement said its Terminus API access for comparison testing was temporary and ended October 15, 2025, at 15:59 UTC. That offer is expired; the announcement does not establish a later direct-API retirement. For a currently documented hosted route, Vercel AI Gateway lists the model as deepseek/deepseek-v3.1-terminus and says it routes requests to an available provider. Confirm the route, provider, and supported features before relying on it. (api-docs.deepseek.com)
What is DeepSeek V3.1-Terminus?
Terminus is an update to DeepSeek V3.1, not a new model family. DeepSeek said the update aimed to reduce Chinese-English mixing and occasional abnormal characters, improve Code Agent and Search Agent performance, and make outputs more stable than the previous version. These are the developer’s stated improvements, not a guarantee that every application or provider will see the same results. (DeepSeek’s Terminus release note)
The official Hugging Face model card also publishes a comparison table for V3.1 and Terminus. For example, it reports SWE Verified scores of 66.0 and 68.4, respectively, and Terminal-bench scores of 31.3 and 36.7. Treat these as model-card benchmark results, not a substitute for evaluating your own tasks; the table alone does not establish how well those results predict your workload. (V3.1-Terminus model card)
Capabilities inherited from V3.1
DeepSeek’s original V3.1 release described hybrid Think and Non-Think inference, a 128K context window, support for the Anthropic API format, and beta strict function calling. These are V3.1-era specifications, not confirmation that a current Terminus API or third-party deployment exposes the same features or limits. Verify behavior with the provider and model identifier you intend to use. (DeepSeek V3.1 release note)
When might Terminus fit your workload?
Terminus may be worth evaluating if you need the V3.1 model line and your tasks depend on consistent multilingual output or agent-style coding and search behavior. The release note and model card give reasons to test those areas, but they do not establish that Terminus is the best choice for your workload—or that it is available through your preferred API.
Use a workload-specific decision:
- Language consistency matters: Include prompts that previously caused language switching, mixed-language output, or unexpected characters. Score correctness and language consistency separately.
- Coding or search agents matter: Test representative tasks with the tools, permissions, and success criteria you use in production. Measure completed tasks and tool-call errors, not just response quality.
- Long context matters: Test with documents at the lengths your application actually sends. Don’t treat the historical 128K V3.1 figure as a current provider limit.
- A production API is mandatory: Confirm Terminus is available and callable through your chosen provider’s current documentation before investing in migration work.
Can you use V3.1-Terminus through a production API now?
DeepSeek’s official documentation records a temporary direct API route for comparison testing, not an ongoing Terminus API commitment. DeepSeek stated that V3.1-Terminus would remain available through a temporary API until October 15, 2025, at 15:59 UTC. Its example used an endpoint URL marked with that expiration date. Both are historical; do not copy that endpoint as current production configuration. The official announcement does not, by itself, say whether every possible direct-provider access path was permanently retired. (api-docs.deepseek.com)
DeepSeek’s current API quick start documents deepseek-flash and deepseek-v4-pro as model choices; it says deepseek-flash is served by DeepSeek-V4.1-Flash-1017. That page does not document a current Terminus selection route, but its model list alone is not a formal statement that Terminus is unavailable in every account or context. (DeepSeek API quick start) (api-docs.deepseek.com)
A currently documented hosted route: Vercel AI Gateway
Vercel AI Gateway currently documents a hosted route for DeepSeek V3.1 Terminus using the model slug deepseek/deepseek-v3.1-terminus. Its page says requests can be made through the AI SDK or through OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages APIs; the gateway authenticates requests and routes them to an available provider. The page lists Novita AI as a provider and displays 131K context and 66K maximum output for that provider listing. Those are gateway/provider-specific details, not evidence of a current direct DeepSeek API route or universal limits. (vercel.com)
Before choosing that route, verify in Vercel’s live documentation and your account:
- The model slug is still available to your account and intended production use.
- Which upstream provider will serve your requests, and what terms apply.
- The context and output limits for that provider and route.
- Whether tool or function calling is supported on that route; don’t assume it from the model’s general capabilities.
- Current pricing, billing units, and any provider-specific conditions.
The Vercel page documents a selectable hosted route, but provider routing, availability, limits, and supported features can differ from DeepSeek’s direct API. (Vercel AI Gateway model page)
If you mean self-hosted weights
DeepSeek’s Hugging Face repository provides V3.1-Terminus weights and displays an MIT license label. Self-hosting is different from using a hosted API: check the repository’s current model instructions, hardware and inference requirements, license terms, and checkpoint limitations before deployment. The model card notes an issue involving the checkpoint’s self_attn.o_proj parameters and UE8M0 FP8 scale format. (V3.1-Terminus model card)
Current DeepSeek API alternatives—not a direct quality comparison
As of November 8, 2026, DeepSeek’s published API pricing page lists the following models and rates. The page states that model prices are per 1 million tokens and displays the rates with a $ symbol. These are not Terminus prices and do not establish that either model matches Terminus on your tasks. Check the live page before budgeting because prices can change. (DeepSeek models and pricing)
| Currently named API model | Version shown by DeepSeek | Peak input, cache miss (per 1M tokens) | Peak output (per 1M tokens) |
|---|---|---|---|
deepseek-flash |
DeepSeek-V4.1-Flash-1017 | $0.30 | $1.20 |
deepseek-v4-pro |
DeepSeek-V4-Pro-0813 | $1.32 | $3.96 |
The pricing page also lists cache-hit input rates and lower off-peak rates. It defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays; all other hours, including weekends and Chinese public holidays, are off-peak. It says off-peak rates are half of peak rates. Confirm the current page and your usage pattern before estimating costs. (DeepSeek models and pricing)
A practical evaluation plan
If a provider confirms it offers Terminus, run a small, controlled evaluation before moving real traffic:
- Build a representative test set. Use real task types—such as multilingual responses, coding tasks, search-assisted work, or long documents—without including sensitive production data unless your policies allow it.
- Set pass criteria in advance. Define what counts as a correct answer, a successful tool call, acceptable language consistency, and acceptable latency or cost.
- Compare like with like. Keep prompts, tool definitions, sampling settings, and test inputs consistent across the models being compared.
- Measure the outcomes that matter. Track task success, tool-call or schema failures, latency, and cost per completed task. Don’t infer production behavior from benchmark scores alone.
- Canary only after the checks pass. Start with a limited, reversible rollout and keep a fallback model available while you monitor failures and cost.
This is a recommended test plan, not a report of tests performed on Terminus.
Bottom line
DeepSeek presented V3.1-Terminus as a V3.1 update focused on language consistency and agent performance, and its model card publishes comparative benchmark scores. DeepSeek’s direct comparison-testing API offer had a stated end date of October 15, 2025; that expired offer does not establish the status of every possible direct-provider route. Vercel AI Gateway currently documents a hosted route with the model slug deepseek/deepseek-v3.1-terminus, while routing through an available provider. Verify provider, limits, tool support, and pricing before selecting it. If you need DeepSeek’s direct API specifically, the current quick start does not document Terminus as a model choice. (api-docs.deepseek.com)
Sources
DeepSeek-V3.1-Terminus | DeepSeek API Docs
DeepSeek-V3.1 Release | DeepSeek API Docs
DeepSeek-V3.1-Terminus model card
chat_python_v3.1_terminus | DeepSeek API Docs
Your First API Call | DeepSeek API Docs