DeepSeek V3.1 Terminus: Reading a Patch Release Honestly

Most patch releases get a line in a changelog and nobody reads further. DeepSeek V3.1 Terminus deserves more attention than that, not because it is a new model, but because its benchmark table is a clean example of what a patch release actually does: it trades. Most numbers go up, a few go down, and whether the update is good for you depends on which column your workload lives in.

DeepSeek released Terminus on 22 September 2025, a month after V3.1. It keeps the same architecture, the same two reasoning modes and the same MIT licence.

What it set out to fix

DeepSeek framed Terminus as a response to issues users reported with V3.1, in two areas:

• Language consistency. V3.1 sometimes mixed Chinese and English in one response and occasionally emitted abnormal characters. Terminus reduces both.
• Agents. Further optimisation of what DeepSeek calls its Code Agent and Search Agent, and an updated template and tool set for the search agent.

The first fix is the one most users noticed. A model that drops into another language halfway through a JSON value or a code comment is not a curiosity in production; it breaks parsers and confuses users. If you saw stray Chinese text or odd characters in V3.1 output, Terminus is the release aimed at your problem.

The benchmark table, read both ways

DeepSeek's model card compares V3.1 and Terminus directly. The gains:

• Humanity's Last Exam — 15.9 to 21.7
• BrowseComp — 30.0 to 38.5
• SimpleQA — 93.4 to 96.8
• SWE-bench Verified — 66.0 to 68.4
• SWE-bench Multilingual — 54.5 to 57.8
• Terminal-bench — 31.3 to 36.7

The flat results: MMLU-Pro 84.8 to 85.0, GPQA-Diamond 80.1 to 80.7, LiveCodeBench 74.8 to 74.9.

And the regressions, which are easy to skim past:

• BrowseComp-zh — 49.2 to 45.0
• Codeforces rating — 2091 to 2046
• Aider-Polyglot — 76.3 to 76.1

The pattern is coherent. Agentic, tool-using work in English improved clearly; raw reasoning barely moved; Chinese-language browsing and competitive programming slipped. A 4.2-point fall on BrowseComp-zh is not noise if Chinese web research is your use case. These are vendor-run numbers, so treat the exact deltas loosely, but the direction of each is the useful signal. Why benchmarks disagree covers how much weight a few points can bear.

A known issue for self-hosters

The model card flags one defect in the checkpoint: the parameters of self_attn.o_proj do not follow the UE8M0 FP8 scale format that V3.1 was otherwise trained in. DeepSeek listed it as a known issue to be fixed in a future release.

This matters mainly if you run the weights yourself on a stack that expects the format consistently. If you call Terminus through an API, the provider has already dealt with how the checkpoint is loaded. If you self-host, test throughput and output quality on your inference engine before you swap it in, rather than assuming it behaves exactly like V3.1 at the kernel level.

Same model underneath

Everything structural carries over from DeepSeek V3.1: 671B total parameters with 37B active, a 128K-token context window, and hybrid thinking and non-thinking modes selected through the chat template. General tool calling still belongs to non-thinking mode, with search handled through its own thinking-mode format, so the agent patterns that worked on V3.1 transfer unchanged.

Terminus is listed on OpenRouter as deepseek/deepseek-v3.1-terminus at $0.27 per million input tokens and $1.00 per million output tokens at the time of writing. That is close enough to V3.1's listed price that cost should not drive the decision either way.

The last of the V3.1 line

The name turned out to be accurate. A week later, on 29 September 2025, DeepSeek released V3.2-Exp, an experimental model that its card says "builds upon V3.1-Terminus" by introducing DeepSeek Sparse Attention. DeepSeek deliberately aligned V3.2-Exp's training configuration with Terminus so that the effect of sparse attention could be measured in isolation.

That gives Terminus a second role: it is the baseline against which the sparse-attention change was judged. If you are evaluating V3.2-Exp, Terminus is the right control, because it differs mainly in attention rather than in training data or post-training.

Should you switch

A short decision rule:

1. Running coding or search agents in English on V3.1: switch. The agent benchmarks all moved up and the language fix removes a real source of broken output.
2. Seeing mixed-language output or stray characters: switch. That is what the release was for.
3. Doing Chinese-language web research or competitive programming: test before switching. Those are the columns that went down.
4. Pinned for reproducibility: stay on V3.1 and call it by its exact id.

Whichever applies, the method is the same: run your own regression set on both versions and compare per task, not on the average. Regression testing a model upgrade shows how to set that up so a 4-point fall on the one category you care about does not hide behind gains everywhere else.