Mistral Small 3.1

tl;dr: Mistral’s 24B multimodal — Apache 2.0, fits a single 4090 or a 32GB Mac, claims 150 tok/s and beats comparable small models like Gemma 3 and GPT-4o Mini at text. Vision landed in this 3.1 release, context jumped to 128K.

What it is

mistral-small3.1 is the Ollama handle for the Mistral Small 3.1 (build 2503) release — the third major drop in the Mistral Small line, with vision and 128K context added on top of Mistral Small 3’s text-only base.

The library page lists vision and tools capabilities plus the 24b size tag, with a 15GB disk footprint and 128K context for both the latest and 24b tags. 786.9K pulls to date. The Mistral README explicitly states Apache 2.0 — direct commercial deploy, no copyleft.

Pull

ollama run mistral-small3.1

Requires Ollama 0.6.5 or higher. Pull the latest Ollama first.

The Ollama page wires it up to Claude Code, OpenCode, Hermes Agent, and OpenClaw via ollama launch.

What the README says

• Lightweight: runs on a single RTX 4090 or a Mac with 32GB RAM.
• ~150 tok/s inference on the recommended hardware.
• Multimodal: text + image input. No audio.
• 128K context window without compromising text performance.
• Function calling for agentic workflows.
• Fine-tunable for specialized domains — Mistral releases both base and instruct checkpoints to support downstream customization.
• Reasoning foundation: the README notes the community (e.g. Nous Research’s DeepHermes 24B) is already building reasoning models on Mistral Small 3 — 3.1 keeps that downstream path open.

Mistral’s own claim: outperforms Gemma 3 and GPT-4o Mini on comparable text benchmarks at this size class.

Sizes on this page

mistral-small3.1:latest and mistral-small3.1:24b both resolve to the 15GB / 128K-context / text+image model. No quantized variants listed on the page — Ollama handles the GGUF conversion.

License

Apache 2.0 per the Ollama library README. Same as the upstream Mistral release — commercial use, modification, redistribution all permitted. No additional rider.

Gotchas

• 4090 floor: 24GB VRAM. Macs need 32GB RAM minimum. The README is explicit about the minimums.
• 15GB on disk for the default. Plan storage if you’re caching multiple Mistral variants.
• Vision is text-prompt + image content; no audio input, no video.
• Mistral ships both base and instruct checkpoints — the Ollama tag is the instruct model. If you need base weights, grab them directly from Mistral.
• Reasoning is community-built. Out of the box it’s a chat/instruct model; DeepHermes-style reasoning requires a separate community finetune.

Related: Ollama for the runtime, Open WebUI for a chat UI, LM Studio for a GUI runner, Gemma 4 as the comparable multimodal small from Google.

Untested here — facts from ollama.com/library/mistral-small3.1 and mistral.ai/news fetched 2026-09-10.