Codestral: 22B Code Model, API Access and License Limits

Mistral released Codestral for more than 80 programming languages with downloadable weights and hosted APIs. See its beta access and non-production license.

A 22B model built for repositories, not snippets

Codestral is a 22 billion parameter model trained on more than 80 programming languages, including Python, Java, C, C++, JavaScript and Bash, with a 32,000 token context window. Mistral AI put the context window ahead of the benchmark tables in its own announcement, comparing it against the 4K to 16K windows on competing code models at the time. [1]

Mistral AI made the weights downloadable on Hugging Face and offered two hosted routes: a free eight week beta on a dedicated endpoint, codestral.mistral.ai, aimed at individual developers testing it inside an editor, and token billed access through its standard API for larger or commercial use. IDE integration shipped through Continue and Tabnine plugins for VS Code and JetBrains, with LlamaIndex and LangChain support for building on top of it. [1][2]

The line the company drew

The weights carry the Mistral AI Non-Production License, a new license written for this release. It permits use for testing, research, personal or evaluation purposes only, and separately bars supplying the model or a derivative of it commercially, including through a hosted or managed service such as a SaaS product or cloud instance, whether or not payment changes hands. Production or commercial use requires a separate license from Mistral AI. [3]

Mistral AI's prior releases, including Mistral 7B, went out under Apache 2.0, and the company had positioned open weights as core to its identity against closed labs. Codestral under a bespoke restrictive license was the news as much as the model: a company that built its name on Apache 2.0 had found a product category, developer tooling with direct enterprise pricing power, where it preferred to keep the commercial door closed and negotiate the key individually. Independent commentary at the time noted the company was careful to call this an "open-weights" release rather than "open source," a distinction Mistral AI did not blur in its own marketing. [3][1]

Sources

  1. Codestral

    Mistral AI · May 29, 2024

  2. Mistral announces Codestral, a code-generation LLM it says is best-in-class

    VentureBeat · May 29, 2024

  3. Codestral: Hello, World!

    Simon Willison · May 30, 2024