Mistral Medium 3.5

What it is

Mistral Medium 3.5 is a large language model from Mistral AI. It appears in the company's API changelog on 28 April 2026, listed as a multimodal model for agentic and coding use, released with open weights ^[1]. The two adjectives carry most of the positioning. "Multimodal" means the model accepts images as well as text. "Agentic" means it is meant to be driven by a program rather than by a person typing one question: a caller hands it a task, the model decides to call a tool, reads what comes back, and keeps going until the job is finished.

Architecture and vision

Medium 3.5 is a dense model with 128 billion parameters. Dense means all of the parameters are used for every token generated, in contrast to mixture-of-experts designs that send each token through only part of the network. Its context window is 256K tokens — the amount of text and images it can hold in view at one time ^[2].

It also carries a vision encoder, and Mistral trained that encoder from scratch rather than attaching one taken from an existing image model. The reason given for building it from scratch is that it had to handle images of variable size and aspect ratio ^[2].

One set of weights, adjustable reasoning

Labs have often shipped separate models for separate jobs: one tuned to follow instructions, one to reason through hard problems, one to write code. Medium 3.5 merges instruction following, reasoning and coding into a single set of weights ^[2], and Wikipedia describes it as Mistral's first flagship model to bring these lines together ^[4]. How much reasoning it does is a per-request setting: callers choose a reasoning effort, exposed in the API as the reasoning_effort parameter, so the same model can answer a quick lookup cheaply or think longer over a hard problem ^[1][2].

Results

Mistral reports 77.6% on SWE-Bench Verified, a benchmark built from real software issues that a model has to resolve, and 91.4 on τ³-Telecom ^[2]. The company said the model has strong agentic capabilities ^[2].

Weights, licence and price

The weights are on Hugging Face under a modified MIT licence, and Mistral says the model can be self-hosted on as few as four GPUs ^[2]. Through the API it costs $1.50 per million input tokens and $7.50 per million output tokens ^[2]. The two routes serve different users: a team that wants the model on its own hardware can download it, and a team that would rather not manage servers can pay per token.

Rollout into Mistral's products

From 22 May 2026 Medium 3.5 became the default model in Le Chat, Mistral's chat product, and replaced Devstral 2 in the Vibe CLI and in the company's remote coding agents ^[2]. Being the default in a chat product decides whether a model is something only developers seek out or something ordinary users meet without choosing it. Devstral 2 had been the model behind the Vibe CLI and the remote coding agents; handing those roles to Medium 3.5 is the concrete form of the "one set of weights" claim, since the coding tool now runs the same model that also follows instructions and reasons.

Retirement

The model's public life was short. Mistral's model overview lists Medium 3.5 as deprecated on 31 October 2026, with retirement scheduled for 30 April 2027, and names Mistral Large 4 as the replacement ^[3]. Deprecation means the model can still be called but is no longer recommended for new work; retirement means it is switched off. From the April 2026 launch to the April 2027 retirement it is in service for about a year, and it was marked deprecated roughly six months after it appeared.

Where it sits

Medium 3.5 is a compact record of several habits in the field at the time it appeared. One is consolidation: instead of shipping an instruction model, a reasoning model and a coding model separately, put them in one set of weights and let the caller dial reasoning effort per request. Another is the open-weights-plus-metered-API arrangement, which lets a small team run the model locally while a product team pays by the token. A third is the emphasis on agentic use — tool calls, remote coding agents, benchmarks that measure whether the model finishes a job rather than whether it writes a plausible paragraph.

The short lifecycle carries its own lesson. Documentation, not reputation, determines whether a given model is the right one to build on: the same model can be the default in a chat product in one year and scheduled for retirement the next, with a named successor already in place.