Mistral AI releases Codestral, its first code model

Background

A code model is a language model trained mainly on source code rather than ordinary prose. Its most common job is completion: reading the code around the cursor in an editor and predicting what comes next, whether that is a single line, a whole function or a block of tests. Two properties shape how useful such a model is in practice — how many programming languages it has been trained on, and how much surrounding code it can keep in view at once.

What was released

On 29 May 2024 Mistral AI released Codestral, a 22-billion-parameter open-weight code model ^[1]. The parameter count refers to the number of weights in the network, a rough measure of model size. Codestral was trained on more than 80 programming languages and has a 32K context window ^[1]. The context window is the amount of text — prompt, surrounding files and generated code together — that the model can attend to in one pass; at 32K tokens it can hold a sizeable slice of a project rather than a single function.

Licence and access

The weights were published under the new Mistral AI Non-Production License, which limits free use to research and testing ^[1][2]. That means use beyond those purposes has to be checked against the licence itself, rather than treated as freely usable software. Alongside the weights, Mistral added a dedicated endpoint at codestral.mistral.ai aimed at IDE completion; it was free during an 8-week beta ^[1]. Integrations at release included Continue.dev, Tabnine, LangChain and LlamaIndex ^[1].

Benchmark result

Mistral reported a score of 73.75 on Kotlin-HumanEval, against 72.05 for GPT-4-Turbo ^[1]. HumanEval is a benchmark built from short programming problems: the model is given a function signature and a docstring, must write the body, and is scored by running unit tests. The Kotlin variant asks for the same task in Kotlin. It is one benchmark in one language, so the number describes Kotlin generation rather than coding in general.

Where it sits

Codestral was Mistral AI's first code model ^[1]. From its release day it was integrated into LlamaIndex and LangChain ^[1], so developers working with those frameworks could call it much as they call other models, without writing a new client. Open weights under a non-production licence, a hosted completion endpoint, and framework integrations available at launch together made up the shape of the release.