Mistral Large

What it is

Mistral Large is a large language model released by the French AI company Mistral AI on 26 February 2024. It was the company's first flagship proprietary model^[1]. The same day, Mistral launched its chat product Le Chat, which ran on Mistral Large; the company also used the model for Le Chat's system-level moderation^[1].

Background and positioning

In its changelog, Mistral described the model — served under the name mistral-large-latest and also known as mistral-large-2402 — as its new flagship model with leading performance^[5]. At launch the company claimed that, among models generally available through an API, it ranked second in the world, after GPT-4^[1].

Claims of that kind are the vendor's own framing rather than an independent measurement, so it is best read as the position Mistral chose for the model: a direct answer to the strongest general-purpose models of the moment, sold as a service rather than released under an open licence. "Proprietary" here means the model is offered commercially; for sensitive use cases Mistral still made self-deployment with the weights available^[1][2].

Technical characteristics

The context window is the amount of text a model can take in at once. Mistral Large shipped with a 32K-token window, and Mistral said it had native fluency in English, French, Spanish, German and Italian^[1] — "native" meaning those languages are covered directly by the model rather than handled through a separate translation step.

It also natively supported function calling and a constrained JSON output mode^[1]. Function calling lets the model do more than produce free text: it can emit a structured request to call an external tool in a predefined format, which a program can then execute. The constrained JSON mode restricts generation to valid JSON so that the output can be parsed directly. At launch, these two features were available only on mistral-small and mistral-large, not on the rest of the line-up^[1].

Mistral Large was released together with Mistral Small (mistral-small-2402), a lower-latency model that Mistral said outperformed Mixtral 8x7B^[1]. Bringing out a large model and a small one on the same day set up a pattern the company has kept since: the large model carries the capability ceiling, the small one carries speed and cost.

How it was offered

Mistral served Mistral Large on La Plateforme, its own developer platform^[1]. On the same day it announced a partnership with Microsoft under which Azure became the model's first distribution partner, with the model available in Azure AI Studio and Azure Machine Learning^[1][2]. For sensitive use cases, self-deployment with the model weights was also offered^[1][2].

Why it mattered

For Mistral, Mistral Large marked a shift in how the company brought a model to market. Instead of a single release, one day carried several things at once: the model on La Plateforme, a distribution partnership with Azure as first partner, and Le Chat as a consumer-facing product running on the same model^[1][2]. Developer-facing capabilities such as function calling and constrained JSON output became part of the product's differentiation from this generation onward^[1].

Later history

The original API version, mistral-large-2402, was deprecated on 30 November 2024 and retired on 16 June 2025^[3]. It was succeeded by Mistral Large 2 in July 2024^[4]. Mistral's model overview now lists Mistral Large 3 as the alternative for users of the retired mistral-large-2402^[3].