Mistral AI releases Mistral Small 3.1 with image input

What Mistral released

On 17 March 2025, Mistral AI released Mistral Small 3.1 under the Apache 2.0 licence^[1]. The licence matters as much as the model. Apache 2.0 is a permissive open-source licence: the weights can be downloaded, inspected, fine-tuned and shipped inside commercial products, with attribution. That is a different proposition from a model reachable only through a paid API, particularly for organisations that cannot hand their data to a third party.

What is new

The release adds image understanding, so the model takes images as input alongside text^[1]. That kind of multimodality is often the first capability dropped when a model is shrunk to run on modest hardware, which is what makes its presence here notable.

Mistral Small 3.1 also comes with a 128K context window^[1]. A context window is the amount of material a model can attend to at one time: the prompt, any retrieved documents, the conversation so far, and its own output as it is generated. A larger window lets long files, whole codebases or questions spanning several documents be handled in a single pass instead of being chopped into chunks that lose their relation to one another.

Performance claims

Mistral said the model outperforms Gemma 3 and GPT-4o mini^[1]. These are the company's own comparisons. Vendor-reported benchmark results are measured on selected tasks, so they are better treated as a starting point for independent evaluation than as a settled ranking.

Where it runs

According to Mistral, Mistral Small 3.1 runs on a single RTX 4090 or on a Mac with 32GB of memory^[1]. Both are machines a developer or a small team may already own, which changes the economics of experimenting: no cluster to reserve, and no per-token bill while iterating. It also means the model can sit on the laptop or workstation where the data already lives.

Where to get it

The model is available on Google Cloud Vertex AI^[1]. Mistral also said it would be available on NVIDIA NIM and Microsoft Azure AI Foundry in the weeks following the release^[1]. Taken together, that puts the same weights in two places at once: downloadable for local use, and served through established cloud platforms for teams that prefer a managed endpoint.

Why it matters

Putting open weights, image input and a long context window into one package is the core of this release^[1]. For a reader trying to follow how small models are developing, the interesting part is not a single benchmark number but the combination: capabilities that recently required a large hosted model are being packed into something that fits on one consumer machine^[1]. Whether it suits a particular job still depends on testing it against that job's own data rather than on the comparison figures published at launch.