DeepSeek V3.1-Terminus: Agent Scores Jump 36%

Introduction

Late on September 22, 2025, DeepSeek quietly shipped V3.1-Terminus. In Latin, Terminus means “the end” — this is almost certainly the closing chapter of the V3.1 line. The next time we meet, it should be V4.

No launch event, no long essay. The official App, DeepSeek web app, mini programs, and DeepSeek API all silently switched to Terminus. The changelog simply said: “Addressed language-consistency feedback and further improved Agent capabilities.”

Anyone who dug into the benchmarks could not stay calm — BrowseComp jumped from 30.0 to 38.5, Terminal-bench from 31.3 to 36.7, with Agent-task gains peaking at 36.5%. This is not a minor patch. It is V3.1’s final metamorphosis.

I. Two Notorious Legacy Bugs Are Finally Gone

Anyone who used V3.1 hit a baffling failure mode: when listing primes, the model would suddenly append “极长” after “2, 3, 5, 7…997” and truncate. Users later traced it to unclean SFT synthetic data containing an array labeled “极长.” Another headache: if your prompt included Go code or version numbers, the model might insert “极” or “extreme” mid-stream and break compilation.

After Terminus shipped, the “极” bug could not be reproduced across repeated attempts. Multilingual mixing (English — or even Russian — appearing inside Chinese prompts) was also dialed down via tokenizer and chat-template fixes. Official figures: Chinese–English mixing down 68%, anomalous character rate down 73%.

II. Agent Capabilities Take Off Across the Board

Here are the official headline numbers:

Benchmark	V3.1 (Aug)	V3.1-Terminus	Delta
MMLU-Pro	84.8	85.0	+0.2
SWE Verified	66.0	68.4	+2.4
SWE-bench Multilingual	54.5	57.8	+3.3
BrowseComp	30.0	38.5	+8.5
Terminal-bench	31.3	36.7	+5.4
Humanity's Last Exam	15.9	21.7	+5.8

Focus on BrowseComp (web browsing) and Terminal-bench (terminal ops) — the core Agent gauges for whether a model can research like a human online and work in a command line. Both rose more than 20%, showing Terminus treated Agent skill as priority one.

On Humanity's Last Exam (HLE), Terminus climbed from 15.9 to 21.7 — enough for a global top-three finish, behind only Grok 4 (25.4) and GPT-5 (25.3), edging Gemini 2.5 Pro (21.6).

Code Agent generation accuracy rose 12%; Search Agent multi-turn retrieval latency improved 40%.

No blind hype, though — community runs show a slight dip on Codeforces contest tasks. If you mainly grind algorithms or LeetCode, classic V3.1 may still feel smoother. For Agent workflows, tool use, and long-horizon tasks, Terminus is the version to pick today.

III. Architecture Unchanged: Same 671B MoE Flavor

Terminus is not a from-scratch retrain — it is a decoder optimization plus template tune on top of V3.1. The stack remains:

• Total parameters: 671B (MoE)
• Active parameters per token: ~37B
• Context window: 128K tokens
• License: MIT (commercial use allowed)

Terminus fully inherits V3.1’s unified chat template — one weight set for both thinking and non-thinking modes. Developers can adopt the latest build seamlessly via the DeepSeek API platform.

IV. How to Start Using DeepSeek V3.1-Terminus

Option 1: DeepSeek web app

Visit chat.deepseek.com — the official App, web client, and mini programs already run DeepSeek-V3.1-Terminus.

Option 2: DeepSeek API platform

Call the latest model through DeepSeek API — the API surface has fully switched to Terminus.

Option 3: Open-source deployment

Open weights are on Hugging Face and ModelScope for customized commercial deployment.

Summary

DeepSeek V3.1-Terminus closes the V3.1 era cleanly: two legacy bugs fixed, Agent skills surging, HLE in the global top three.

Whether you try the DeepSeek web app or integrate via the DeepSeek API platform, this upgrade is worth a hands-on test. Next time we meet, it should be DeepSeek V4.

Further Reading

• DeepSeek V4 Pro Official Release
• DeepSeek V4 Official Launch on July 15
• DeepSeek V4 API Docs