Codestral

What Codestral is

Codestral is Mistral AI's first model built specifically for code. The first version, released on 29 May 2024, is an open-weight model with 22 billion parameters, trained on more than 80 programming languages^[1]. It shipped with a 32K-token context window, which Mistral contrasted with the 4K–16K windows of competing code models at the time^[1]. Context length matters for code because a model can only work with what fits in the window: a long function, several related files, or a configuration file together with the code that reads it.

Licence and access

The weights were published under the Mistral AI Non-Production License (MNPL), which covers research and testing; commercial licences are available on request^[1]. Developers could reach the model in two ways: a dedicated endpoint at codestral.mistral.ai for IDE instruct and fill-in-the-middle use, free during an eight-week beta, and the general api.mistral.ai^[1]. Fill-in-the-middle (FIM) means the model is given the code on both sides of a gap and asked to write what belongs in the gap — the pattern that editor autocomplete relies on.

Launch integrations and benchmark

Codestral was integrated with LlamaIndex, LangChain, Continue.dev and Tabnine from its release day^[1]. On Kotlin-HumanEval, a researcher at JetBrains measured 73.75 for Codestral against 72.05 for GPT-4-Turbo^[1]; Mistral presented the result as Codestral surpassing GPT-4-Turbo's score^[1]. HumanEval is a standard set of small programming problems used to compare code models, and the Kotlin variant poses the same kind of task in Kotlin rather than Python.

Codestral 25.01

On 13 January 2025 Mistral released Codestral 25.01, which extended the context window to 256K tokens and generated code about twice as fast^[2]. It scored 71.4% on average on HumanEval, up from 65.6%, and debuted at number one on the LMsys Copilot Arena^[2].

Codestral 25.08 and the coding stack

Codestral 25.08 followed on 30 July 2025; Mistral described it as its cutting-edge language model for code completion^[4]. Compared with earlier versions, it increased accepted completions by 30% and retained code by 10%, with 50% fewer runaway generations^[3] — completions that keep producing output instead of stopping at a sensible point. It shipped as part of an enterprise coding stack that also includes Codestral Embed, Devstral and the Mistral Code IDE plugin^[3].

Model lifecycle

The original release was listed with a deprecation date of 2 December 2024, with Codestral named as the alternative in Mistral's model overview^[4]. codestral-2405 was retired on 16 June 2025 and codestral-2501 on 30 November 2025; Codestral 25.08 remains active as a premier model^[4].

Place in the line

Read across the releases, Codestral began as a single open-weight model meant to be dropped into an editor, and later versions worked on the things that matter in day-to-day coding: how much code fits in the context window, how quickly completions come back, and how often a suggestion is accepted and then kept. The last of these releases also placed the model inside a wider enterprise offering rather than shipping it on its own^[4][3].