Mistral Medium 3

Mistral Medium 3 is positioned as a new middle tier in the Mistral lineup, designed to deliver frontier-class results without the price tag of the largest models. It is a multimodal model that accepts both text and image inputs, so it can read long documents and interpret visuals alongside natural language. The model targets professional workloads where coding ability and multimodal understanding matter most, and it is engineered to be deployed either as a hosted service or inside an enterprise's own infrastructure. Its design also supports hybrid and on-premises setups, custom post-training, and integration with existing enterprise tooling, which is why IBM describes it as runnable on a modest footprint of just four H100 GPUs for organizations that want a manageable on-premise footprint rather than a sprawling cluster.

The model is part of Mistral's broader effort to compress the performance of much heavier systems into a more efficient package, with the launch framing it as reaching around ninety percent of a top-tier competitor on common benchmarks while costing roughly an order of magnitude less to run. It leads in professional areas such as coding and multimodal reasoning, and surpasses leading open models like Llama 4 Maverick as well as enterprise peers such as Cohere Command A on the evaluations Mistral highlights. Practical strengths show up in long-context document analysis, agent-style workflows where the model calls external tools, and low-latency applications that benefit from its efficient footprint. With availability across multiple enterprise platforms and support for custom post-training, it fits well for teams that want a capable, controllable model they can tailor and host on their own terms.