Mistral AI releases open-weight Mistral Medium 3.5

Background

Mistral Medium 3.5 is the mid-sized entry in Mistral AI's model line, released as an open-weight model: the trained parameters are published so that others can download, run and modify them, rather than being reachable only through an API. For coding work this has a practical consequence — a team that cannot send source code to a third-party service can still run the model on its own hardware.

Architecturally the model is dense rather than a mixture of experts. In a dense transformer every parameter takes part in the computation for every token; a mixture-of-experts model activates only a fraction of its parameters per token. Dense designs are simpler to serve and simpler to reason about, at the cost of more compute per token at the same parameter count.

What was released

Mistral's changelog dates the release of Mistral Medium 3.5 to 28 April 2026^[1]. The model is dense, has 128B parameters and a 256K context window, and accepts both text and images as input^[2]. Its weights were published under a modified MIT licence^[2].

A 256K context window means a very large amount of material — many files from a repository, or a long document — can be supplied in a single call instead of being fed in piece by piece. The licence is a modification of the permissive MIT licence, the terms under which the weights may be used, changed and redistributed.

Results and design choices

Mistral reports 77.6% on SWE-Bench Verified^[2], a benchmark built from real software issues drawn from open-source projects: the model has to produce a patch, and it is scored on whether the project's tests then pass. Mistral also reports a score of 91.4 on τ³-Telecom^[2].

For images, Mistral says the vision encoder was trained from scratch, and specifically for variable image sizes and aspect ratios^[2], rather than requiring every picture to be resized to one fixed shape. Reasoning is adjustable at inference time through a reasoning_effort parameter^[1]: the same model can be asked to think for longer or shorter before it answers, trading latency against accuracy.

Deployment

On 22 May 2026 the model began powering Vibe's remote coding agents and became the default model in Le Chat, Mistral's chat product^[2]. The two uses point in different directions. In Vibe, the model is the engine inside an agent that works on code remotely, where a long context and a strong result on a software-engineering benchmark are the qualities that matter. As the default in Le Chat, it becomes the model an ordinary user meets first, where image input and the ability to dial reasoning effort up or down are the more visible features.

Why it matters

The release is a data point about how open weights and frontier-adjacent capability are being combined. A 128B dense model is large enough that self-hosting it is a real infrastructure commitment, but publishing the weights under a modified MIT licence gives users an option that a closed API does not: the ability to inspect, fine-tune and keep running the model on their own terms. The adjustable reasoning parameter reflects a broader pattern in which a single model is expected to serve both fast, cheap responses and slower, more deliberate ones.