DeepSeek-V3.1

tl;dr: DeepSeek-V3.1-Terminus — one 671B MoE model that flips between thinking and non-thinking modes via the chat template, with sharper tool calling and faster reasoning than DeepSeek-R1-0528 at comparable quality.

What it is

deepseek-v3.1 is the Ollama-side handle for the V3.1-Terminus release. Same MoE skeleton as V3, retuned for agentic workloads. The “hybrid” pitch: keep one set of weights, switch the chat template, get a non-thinking chat model or a slow-cooked reasoning model.

The library page lists tools and thinking capabilities plus the 671b size tag, with 404GB on disk for the latest and 671b tags and a 160K context window. 726.5K pulls to date, no separate “thinking” or “non-thinking” tags on the Ollama side — the toggle lives in the prompt template.

Pull

ollama run deepseek-v3.1
# or pin the explicit MoE weight set:
ollama run deepseek-v3.1:671b

Applications the Ollama page wires up via ollama launch: Claude Code, OpenCode, Hermes Agent, OpenClaw.

What the README says

The Ollama README frames the Terminus update against V3.1 as three small wins:

• Language consistency — fewer CN/EN mix-ups, no random characters mid-stream.
• Agent upgrades — stronger Code Agent and Search Agent performance from post-training.
• Hybrid thinking mode — chat template switches between thinking and non-thinking; one set of weights covers both call patterns.
• Smarter tool calling — measurable lift in tool-use and agent benchmarks.
• Higher thinking efficiency — V3.1-Think reaches comparable answer quality to DeepSeek-R1-0528 with less wall time.

Trade-off: this is a thinking model where the chain-of-thought is exposed when you ask for it. If you want raw text, drop the template flag.

Sizes on this page

The library shows deepseek-v3.1:latest and deepseek-v3.1:671b as the same 404GB / 160K-context / text-only weight set. The page header advertises seven tags in total — distilled smaller variants also exist (see the page for the current tag list), but only the headliner is shown inline here.

License

The Ollama library page does not state a license for this model. DeepSeek-V3.1 ships under the DeepSeek License, which permits commercial use but adds use-based restrictions (no use for military, ideological manipulation, etc.). Confirm against the upstream repo before shipping commercial output.

Gotchas

• 404GB on disk for the 671B MoE. Plan storage before pulling, and expect multi-GPU or a Mac Ultra to actually run it.
• Hybrid thinking means the same model produces verbose reasoning blocks when configured for thinking. Disable the thinking template if you want pure chat output.
• Distilled smaller variants inherit the upstream license family — always check before commercial deployment.
• 671B is text-only. V3.1-Terminus does not add the multimodal extensions later DeepSeek-VL lines ship.

Related: DeepSeek-R1 for the dedicated reasoning family, GPT-OSS for the OpenAI open-weight counterpart, Ollama for the runtime.

Untested here — facts from ollama.com/library/deepseek-v3.1 fetched 2026-09-10.