Mistral AI launches the Le Chat assistant

Background

An API is the interface through which one program talks to a model: a developer sends text in and receives text back, without any window or buttons. Everything a company ships through an API is aimed at people who write code. A chat interface points the other way — a page where a person types a message and reads a reply, with the model itself kept behind a product name.

Mistral AI's API changelog is where the company records what it has added to that developer-facing surface ^[1].

What happened

On 26 February 2024, the changelog listed the launch of Le Chat together with two new models, Mistral Large and Mistral Small ^[1]. The entry introduced Le Chat as "the brand new chat interface Le Chat to easily interact with Mistral models" ^[1].

So the same announcement carried a product for end users and models for developers. Someone opening Le Chat and someone calling the API were, in that sense, meeting the same underlying models through two different doors.

Key details

The changelog also brought function calling and JSON mode to the API on the same day, available for both Mistral Small and Mistral Large ^[1]. Function calling lets a model ask for a specific tool to be run, naming the function and supplying arguments in a structured form instead of writing the request out as prose; JSON mode constrains a reply to valid JSON, the format programs can parse without guessing at the shape of the text. Both features matter for the same reason: they turn a model's output into something a piece of software can act on reliably.

The same entry introduced `mistral-small-2402` as a new model ^[1], giving the smaller of the two a versioned name in the API.

Mistral described Le Chat as a beta "assistant demonstrator" that exposes its models, including Mistral Large ^[2]. The word demonstrator is doing real work there: Le Chat was presented not as a finished consumer product but as a way to show what the models could do in conversation.

One further detail concerns how the interface itself was run. Mistral used Mistral Large to build Le Chat's system-level moderation ^[2]. Moderation here sits at the level of the system rather than inside any single answer: a filter layer around the assistant, and in this case one built with the company's own largest model.

Why it matters

A single changelog entry covered two audiences at once. Developers received structured-output tooling — function calling and JSON mode — of the kind needed to wire a model into an application, plus a named small model to use it with ^[1]. Everyone else received a chat window ^[1].

Using Mistral Large for moderation is a small but telling design choice: the same class of model that answers questions was also put to work policing the interface around them ^[2]. And calling Le Chat a beta assistant demonstrator sets the expectation for what it was at launch — a way to interact with the models, not a claim that the product was finished ^[2].