Mistral Small 3.1 24B: Vision and 128K in an Open Model

Mistral Small 3.1 is the middle child of a three-release run, and middle children get skipped. Small 3 set the size and the licence; Small 3.2 polished the behaviour. But 3.1 is where the two capabilities that define the line arrived: image input and a long context window. Understanding it explains why the family looks the way it does.

It also has one surprising property — on a few vision benchmarks it scores slightly above its own successor — which is the only good reason to deliberately pin it today.

What changed from Small 3

Mistral released Small 3.1 on 17 March 2025, under Apache 2.0, with both an instruct and a pretrained base checkpoint on Hugging Face. The size stayed at 24B parameters, and the headline performance claim of around 150 tokens per second carried over from Mistral Small 3.

Two things changed, and both are large:

• Context went from 32K to 128K tokens. That moves the model from short-prompt work into whole documents, long transcripts and multi-file code context.
• It accepts images. The announcement describes improved image understanding and lists document verification, visual quality inspection and image-based customer support as target uses.

Mistral kept the deployment story intact: it says the model runs on a single RTX 4090 or a Mac with 32 GB of RAM, which in practice means quantised. Adding vision without growing the parameter count is what made it interesting — you got a multimodal model you could still run on one workstation.

The claim, and how to read it

Mistral called it "the best model in its weight class" and said it outperforms comparable small models including Gemma 3 and GPT-4o mini across text, multimodal, multilingual and long-context tasks. The detailed scores in the launch post are in charts rather than text.

The Small 3.2 model card is more useful here, because it lists 3.1's numbers next to 3.2's on the same harness. For 3.1 it gives MMLU 80.62%, MMLU-Pro 66.76%, MATH 69.30%, HumanEval Plus (pass@5) 88.99%, MMMU 64.00%, MathVista 68.91%, ChartQA 86.24% and DocVQA 94.08%. Those are vendor-run figures, but at least they are like for like.

A "best in weight class" claim is always conditional on which rivals were chosen and when. Any 2025 comparison against Gemma 3 or GPT-4o mini says little about how 3.1 fares against what is available in late 2026. Launch-day tables are a starting point for a shortlist, not a verdict.

Testing the vision side properly

If image input is why you are considering 3.1, test it on images that look like your production traffic, not demo photos. A useful minimum set:

1. Dense documents. Scanned invoices or forms where a single misread digit matters. DocVQA-style scores are high for this model, but your scans have their own noise.
2. Charts. Ask for specific values, not a description. A model that summarises a chart well can still misread the axis.
3. Screenshots with small text. Image resolution handling differs across serving stacks, so the same model can behave differently on two hosts.
4. Negative cases. Images that do not contain the requested information. You want a refusal, not a plausible invention.

Vision-language models compared goes further on what to test before committing, including which jobs images genuinely help with.

Small 3.1 or Small 3.2

Three months later Mistral shipped Small 3.2 as a direct update. On Mistral's own comparison, 3.2 is clearly better at instruction following and chat quality — WildBench v2 rises from 55.6% to 65.33% and Arena Hard v2 from 19.56% to 43.1% — and runaway, repetitive generations fall from 2.11% to 1.29% on Mistral's measure. Function calling also became more robust.

The exception is vision. The same table shows 3.1 slightly ahead on MMMU (64.00% against 62.50%), MathVista (68.91% against 67.09%) and AI2D (93.72% against 92.91%), while 3.2 edges ahead on ChartQA and DocVQA. These are differences of one or two points and could easily vanish on your data, but they are the only measured reason to prefer the older model.

The rule: choose 3.2 by default. Stay on 3.1 only if you already run it on an image-heavy pipeline and a side-by-side test on your own images shows 3.2 doing worse.

Availability and price

Mistral's documentation lists Small 3.1 as deprecated on 6 November 2025 and retired on 30 November 2025, with Mistral Small 4 as the suggested replacement. The open weights remain available, and some third parties still host it.

At the time of writing it is listed on OpenRouter at about $0.35 per million input tokens and $0.56 per million output — noticeably more than the listing for the newer Small 3.2. Aggregator prices for retired models reflect which hosts still bother to serve them rather than what the model costs to run, which is another argument for moving to 3.2 or self-hosting. If you are pinned to this snapshot for reproducibility, keep a copy of the weights you can serve yourself, because no host is obliged to keep a retired model online.

For a new project that wants a hosted Mistral endpoint with vision, the only current option on Mistral's own API is Mistral Small 4, which is a different and much larger model.