Mistral Small 3.1

What it is

Mistral Small 3.1 is a language model released by Mistral AI on 17 March 2025 under the Apache 2.0 licence. Its checkpoint names indicate 24 billion parameters, and the official Hugging Face model card states the same figure.^[1][4] That places it in the "small" tier of models: systems built to run on hardware a developer or a small team might already own, rather than on a large cluster of accelerators.

Where it came from

The model is an extension of Mistral Small 3. Its main addition is image understanding, meaning it can take pictures as input alongside text — something the earlier Small model could not do.^[1] At the same time, Mistral extended the context window to 128K tokens, so the model can work through a much longer document in a single pass.^[1]

Two checkpoints were published. One is a base model, trained on text and images without further alignment; the other is an instruction-finetuned version built on top of it. The Hugging Face model card identifies the instruction checkpoint as a fine-tuned version of Mistral-Small-3.1-24B-Base-2503.^[4]

What it can do

Mistral reported that the model runs at 150 tokens per second, and claimed that it outperforms Gemma 3 and GPT-4o mini.^[1] Those are the company's own comparisons rather than an independent benchmark, so they are best read as a statement of where Mistral positioned the model in the market.

Just as important for who could actually use it is the hardware requirement: the model runs on a single RTX 4090, or on a Mac with 32GB of RAM.^[1] A model in that class can be run locally, on a workstation or a laptop-class machine, without a cloud account — which matters for privacy-sensitive work, for offline settings, and for developers who want to iterate cheaply.

The combination of image input, a long context window and modest hardware demands is what defined the release. A long context makes it practical to hand the model a whole report, contract or codebase file; image understanding widens the range of documents it can accept to include scans, screenshots and charts.

How it was distributed

Mistral published base and instruction checkpoints on Hugging Face, and made the model available on its own La Plateforme and on Google Cloud Vertex AI at launch.^[1] The company also said the model would be available on NVIDIA NIM and Microsoft Azure AI Foundry in the coming weeks.^[1] The Apache 2.0 licence allowed others to download, modify and deploy the weights commercially without negotiating a separate agreement.

Where it sits in the line

The model was a step in a fast-moving series rather than a one-off release. Mistral Small 3.2, published under the checkpoint name mistral-small-2506, followed on 20 June 2025.^[2] Small 3.1 was then retired on 30 November 2025, and Small 3.2 on 31 July 2026; both were replaced by Mistral Small 4.^[3]

The short life of the model says something about the pace of the small-model tier at the time: capable enough to run on a single graphics card, but superseded within months. Its lasting contribution was to show that image understanding and a 128K context window could be packaged into a model that fits on a desktop machine, and that such a model could be released openly under a permissive licence.