Short answer: DeepSeek-V3.1 is a hybrid reasoning model with thinking and non-thinking modes. DeepSeek later announced V3.1-Terminus as an update to V3.1. The current DeepSeek API documentation lists deepseek-flash and deepseek-v4-pro; it says deepseek-flash is served by DeepSeek-V4.1-Flash-1017. That does not, by itself, establish whether V3.1 or Terminus is available through a particular account or another host. If your application needs that exact version, confirm the model revision and identifier with the service you plan to use. (DeepSeek API model and pricing documentation)
What DeepSeek-V3.1 is
Together AI describes DeepSeek-V3.1 as a hybrid model with thinking and non-thinking modes. Its page lists 671 billion total parameters, 37 billion activated parameters, and a 128K-token context window. These are specifications reported by Together AI; a hosted API may set its own limits or expose a different model revision. (Together AI’s DeepSeek-V3.1 page)
DeepSeek’s September 21, 2025 announcement said “DeepSeek-V3.1 is now DeepSeek-V3.1-Terminus.” It described the update as addressing language consistency and improving Code Agent and Search Agent performance, and said it was available through the app, web product, and API at that time. Those statements describe the announcement and release-time availability; they do not establish what a specific API account or host serves now. (DeepSeek’s Terminus announcement)
The open-weight model and a production API are different selection paths. If you intend to self-host or use a host, confirm the exact model revision it offers rather than assuming that a direct API identifier selects the same checkpoint. (DeepSeek’s Terminus announcement)
Where its capabilities may fit
DeepSeek presents V3.1 as supporting reasoning, coding, and agent tasks. Math or research analysis, debugging, code review, and multi-step tool workflows are plausible evaluation targets—not guaranteed production outcomes. (DeepSeek-V3.1 release announcement)
Together AI’s model page reports these mode-specific results:
| Benchmark | Non-thinking | Thinking |
|---|---|---|
| MMLU-Redux | 91.8% | 93.7% |
| LiveCodeBench Pass@1 | 56.4% | 74.8% |
| AIME 2024 Pass@1 | 66.3% | 93.1% |
These are Together AI’s reported figures, not results independently validated here. The page does not provide evaluation dates or enough methodology to treat them as a like-for-like comparison with another model. Use the scores as a reason to test relevant tasks, not as a production-quality or speed guarantee. (Together AI’s DeepSeek-V3.1 page)
DeepSeek’s original V3.1 API announcement mapped deepseek-chat to non-thinking mode and deepseek-reasoner to thinking mode. That is historical release guidance; it does not confirm that those identifiers still route to V3.1 today. (DeepSeek-V3.1 release announcement)
Production API: check the model name before you integrate
DeepSeek’s current API documentation lists deepseek-flash and deepseek-v4-pro as model identifiers, and gives OpenAI-format and Anthropic-format base URLs. It identifies deepseek-flash as served by DeepSeek-V4.1-Flash-1017. The documented model-list endpoint provides currently available model identifiers and metadata. These documents establish the identifiers and routing they describe; they do not establish that every legacy V3.1 identifier is unavailable across all accounts or providers. If you need V3.1 specifically, confirm its availability and exact identifier with DeepSeek or your chosen host before integration. (DeepSeek API model and pricing documentation; model-list endpoint documentation)
A practical pre-integration check:
- Decide whether you need the exact V3.1/Terminus weights or simply a current DeepSeek API model.
- For direct API use, check the current model list and ask support if the exact legacy version is not clear. Do not assume that a newer identifier selects V3.1. (DeepSeek model-list endpoint documentation)
- If using a host, confirm the exact V3.1 revision and identifier, mode controls, tool-calling support, context and output limits, rate limits, regional availability, and data-handling terms. These are provider-specific details.
- Run a staging test against the exact identifier you plan to deploy. Verify the returned model behavior and measure your workload’s quality, latency, errors, and cost before switching production traffic.
Together AI says DeepSeek-V3.1 is available on dedicated endpoints and directs users to contact the provider for pricing. Its page does not give a numeric V3.1 price or billing unit. Confirm current availability and written pricing for the particular endpoint; do not infer V3.1’s price from prices for other models on the same site. (Together AI’s DeepSeek-V3.1 page)
How to decide whether it belongs in your stack
Consider V3.1/Terminus when you have a concrete reason to test that version—for example, a workload that may benefit from its two modes or a requirement to run its open weights—and can access a provider or deployment that identifies the exact revision. If you mainly need a current direct DeepSeek API integration, check the current model list and confirm whether its listed models meet your needs; the current deepseek-flash mapping points to V4.1-Flash-1017, not V3.1. (DeepSeek API model and pricing documentation)
For a useful comparison, run candidate models against the same representative tasks. For example, use held-out code-review requests and score correctness, useful findings, false positives, and adherence to your required format. If the task involves tool use, record successful tool calls and task completion—not just prose quality. Measure latency and cost on the same workload, and confirm the provider’s billing unit and operating limits. This is a proposed evaluation method, not a reported test result for V3.1.
Bottom line
DeepSeek announced Terminus as an update to V3.1, but the current API documentation’s deepseek-flash identifier maps to V4.1-Flash-1017. That alone does not prove whether a legacy V3.1 revision is available through a particular account or host. Before adopting it, verify the exact model ID and revision, test your own workload, and confirm applicable limits and price. (DeepSeek’s Terminus announcement; current DeepSeek API model and pricing documentation)
Sources
Together AI: DeepSeek-V3.1 model page
DeepSeek-V3.1 is now DeepSeek-V3.1-Terminus