Mistral Open-Source Models

Mistral offers models at every scale, from the lightest (3B parameters, runnable on a phone) to the most powerful (675B, requiring a data center). Here are the most relevant models for healthcare use.

Small models (edge / workstation)

Medium models (dedicated server)

Frontier model

The major advantage of open-weight models is being able to run them on your own hardware. No data leaves your infrastructure. Here are recommendations based on your equipment.

As with GPT-OSS, Mistral's models are more precisely open-weight than open-source in the strict sense: the trained weights and inference code are public, but the training data and alignment procedures are not.

For clinicians, this means: you control where the model runs and who accesses the data, but you cannot verify what it was trained on. Mistral's European advantage (francophone culture, French team) makes it plausible that French and European contexts are better represented in the training data — but this is an inference, not a certainty.

Mistral's open-weight ecosystem is the most comprehensive available in Europe. It combines three advantages that neither OpenAI (GPT-OSS) nor Meta (LLaMA) offer simultaneously: a European company, native multilingualism centered on French, and models at every scale.

For a French mental health facility, the most realistic short-term scenario is: Mistral Small 3.1 on an internal server (or certified health data host) for writing assistance and clinical exploration, complemented by Le Chat Enterprise for uses requiring the frontier model.

The release of Mistral Large 3 under Apache 2.0 (December 2025) is a strong strategic signal: for the first time, a European frontier model of 675 billion parameters is freely downloadable and usable by any institution. Even though running it requires a data center, this opens the way to sovereign deployments at the scale of a hospital group (GHT — Groupement Hospitalier de Territoire) or a regional health agency (ARS — Agence Régionale de Santé).