What is DeepSeek-V3.1-Terminus?
DeepSeek-V3.1-Terminus is an open-weight large language model provided by DeepSeek. It is not a completely separate model family: it is a revised checkpoint in the DeepSeek-V3.1 line. The Terminus update concentrates on reliability and practical agent performance, especially in situations where a model must maintain consistent language, generate code, search through information, or complete a sequence of tool-assisted steps.
The model has two operating styles. Non-thinking mode is intended for lower-latency conversation, instruction following, and ordinary text generation. Thinking mode provides a more deliberate reasoning-oriented generation path for tasks that benefit from additional planning. This makes the checkpoint useful when one deployment needs both faster responses and deeper problem solving without switching to an entirely different model.
DeepSeek released the weights under the MIT License. That permissive license supports research, modification, and commercial deployment subject to the license terms, but it does not remove the operational work involved in running a model of this size.
Where it fits in DeepSeek's lineup
DeepSeek-V3.1-Terminus belongs to the V3.1 generation and should be understood as a refinement rather than a successor with a new architecture. Compared with the earlier DeepSeek-V3.1 checkpoint, DeepSeek described Terminus as reducing Chinese-English language mixing and abnormal-character problems while improving code-agent, search-agent, and long-horizon task behavior.
That positioning matters because the model is aimed less at a simple hosted chat experience and more at developers evaluating open models for their own infrastructure. DeepSeek's consumer services and other API offerings are separate products. The Terminus weights can be deployed independently, while its dedicated first-party comparison endpoint was explicitly temporary and is no longer a current access route.
Capabilities and technical specifications
The verified context window is 128,000 tokens. A token is a small unit of text used by language models, so this capacity can accommodate long documents, substantial codebases, extended conversations, or multi-step agent state, although the practical limit also depends on the serving software and available memory. No authoritative maximum output-token limit is identified in the supplied model documentation.
The underlying V3 architecture is a mixture-of-experts, or MoE, system. It has approximately 671 billion total parameters, with about 37 billion activated per token. In simple terms, the model contains a very large collection of learned components, but only a portion is used for each piece of generated text. This can improve computational efficiency relative to activating every parameter at every step, but the full checkpoint is still extremely demanding to host.
DeepSeek distributes the checkpoint in several tensor formats, including BF16, FP8, and F32-related files. Compatible deployment routes include Transformers, vLLM, SGLang, and similar inference systems. Quantization and distributed inference may reduce practical memory requirements, but the model is not a realistic ordinary single-GPU installation without substantial optimization.
Reasoning, coding, and agent performance
Thinking mode makes Terminus suitable for tasks that require planning, decomposition, or checking intermediate steps. It can be used for mathematical and analytical prompts, complex instruction following, code generation, and workflows in which the model proposes actions for an external tool runner. Non-thinking mode is the more appropriate choice when response time and throughput matter more than extended reasoning.
DeepSeek's published comparison reported improvements over DeepSeek-V3.1 on several evaluations. SWE-bench Verified increased from 66.0 to 68.4, SWE-bench Multilingual from 54.5 to 57.8, and Terminal-Bench from 31.3 to 36.7. The reported BrowseComp score rose from 30.0 to 38.5, while SimpleQA increased from 93.4 to 96.8. These are provider-reported comparison results, not guarantees for a particular application. Real-world performance depends on prompts, tools, retrieval data, inference settings, and the quality of application-level validation.
The results are most relevant to software engineering assistants, terminal-oriented automation, search-agent prototypes, and long-running workflows. The model can produce tool or function calls under an orchestration layer, but the base checkpoint does not independently browse the web, run shell commands, or execute arbitrary tools. A separate application must interpret the model's requested action, perform it safely, and return the result.
Supported inputs and outputs
Terminus is primarily a text model. It accepts text input and produces text output, including reasoning-oriented text and structured tool-call requests where the serving interface supports them. It does not provide native image, audio, video, music, embedding, or speech output according to the supplied specifications.
The model's tool-use capability should therefore be distinguished from native multimodal or autonomous-agent capability. A developer can connect it to search, terminal, database, or application tools, but those tools are external. Similarly, the model's structured-output or JSON support depends on the compatible API or inference stack used to serve it; developers should validate generated data rather than assuming that every deployment enforces a schema identically.
Deployment and API status
The most important practical distinction is that the open-weight checkpoint remains available while the dedicated first-party Terminus API endpoint does not. DeepSeek provided a temporary route named with v3.1_terminus_expires_on_20251015 for comparison testing. It was scheduled to end on October 15, 2025, at 15:59 UTC, and the supplied research identifies that endpoint as retired.
Users evaluating Terminus today should plan around self-hosting or a compatible third-party inference provider rather than assuming that the old DeepSeek endpoint can be used for production traffic. Self-hosting provides more control over data handling, deployment configuration, and model versioning, but requires distributed infrastructure, memory planning, monitoring, and safety controls. Hosted alternatives may be simpler, but their availability, pricing, and supported features must be verified separately.
Pricing and license
There is no current per-token fee from DeepSeek for downloading and self-hosting the MIT-licensed weights. Self-hosting is not free in an operational sense: hardware, cloud GPU rental, storage, networking, engineering, and maintenance can become the dominant costs.
During the temporary first-party API period, reported pricing was approximately $0.56 per million input tokens for cache misses, $0.07 per million cached input tokens, and $1.68 per million output tokens. These figures are historical and should not be interpreted as current API pricing or evidence that the retired endpoint is available. Any third-party hosted deployment may use a different price, context limit, throughput policy, or licensing arrangement.
Main limitations
- Very large deployment footprint: the approximately 671B-parameter checkpoint requires substantial memory and generally calls for quantization, model sharding, or distributed inference.
- No current dedicated first-party endpoint: the Terminus-specific DeepSeek API route was temporary and ended on October 15, 2025.
- External tools are required: web search, terminal execution, retrieval, and other actions must be implemented by an application around the model.
- Known FP8 documentation issue: the checkpoint documentation identifies a scale-data-format issue affecting
self_attn.o_proj, which DeepSeek said would be corrected in a future release. - Benchmark results are not guarantees: provider-reported evaluation gains may not translate directly to every coding, search, or business workflow.
- Text-only model operation: it is not a native image, audio, or video generation system.
Speed, cost, and capability trade-offs
Terminus offers a useful choice between faster non-thinking generation and more deliberate thinking generation, but neither mode removes the cost of serving a very large model. Compared with smaller hosted language models, it may be harder and more expensive to deploy, even when its open license reduces per-token vendor fees.
The model is most attractive when control over weights, long context, coding quality, reasoning behavior, or agent experimentation matters enough to justify infrastructure complexity. A smaller hosted model may be more appropriate for high-volume classification, simple support replies, low-latency interactions, or teams without GPU and inference expertise. A managed API may also be preferable when predictable uptime and operational simplicity are more important than self-hosting control.
When to choose DeepSeek-V3.1-Terminus
Choose DeepSeek-V3.1-Terminus when you need an open-weight model for self-hosted experimentation or production evaluation and can support its infrastructure requirements. It is a strong candidate for coding assistants, software-engineering analysis, long-context document processing, multilingual text generation, search-agent prototypes, terminal-oriented automation, and research into hybrid reasoning workflows.
Its thinking and non-thinking modes are especially useful when the same application needs to balance response speed against more careful reasoning. The MIT license can also be valuable to organizations that want to inspect and control the deployed checkpoint rather than depend entirely on a closed hosted service.
Consider another option when you need a turnkey API, guaranteed access to the retired Terminus endpoint, native image or audio capabilities, autonomous tool execution, or deployment on modest hardware. Terminus is best viewed as a capable open checkpoint that requires engineering around it, not as a complete hosted agent platform.