Mistral's first code-specialist AI is here
Mistral AI just dropped Codestral, its first-ever generative model fine-tuned specifically for coding, sporting 22B parameters and a 32k context window. Fluent in over 80 programming languages, it shines at fill-in-the-middle completions and repo-level tasks. Why it matters: Developers get an open-weight, low-latency engine directly inside their IDEs, though commercial deployments require a custom license.
Codestral is Mistral AI's first dedicated generative model for coding. Sized at 22B parameters, it is built to help developers autocomplete functions, fill in code blocks, generate tests, and chat about code.
The model is fluent in more than 80 programming languages. Alongside staples like Python, Java, C, C++, JavaScript, and Bash, it handles more specialized languages such as Swift and Fortran.
It features a 32k context window, which is significantly larger than the 4k to 16k limits found in many older coding models. That gives it enough headroom to parse extensive multi-file repositories in one go.
Crucially, Codestral supports fill-in-the-middle (FIM) completion right out of the box. Instead of merely continuing code downward, it inspects what comes before and after your cursor to insert precisely what is missing.
Benchmarks and IDE integrations
Codestral posted strong results across standard code evals, including RepoBench EM for long-range repository completion, HumanEval pass@1, MBPP, and Spider for SQL.
External feedback has been positive. JetBrains researcher Mikhail Evtikhiev noted that on their Kotlin-HumanEval benchmark (at T=0.2), Codestral reached a 73.75 score, topping GPT-4-Turbo at 72.05 and GPT-3.5-Turbo at 54.66.
Developer tools jumped on it immediately. Continue.dev and Tabnine integrated Codestral into their VS Code and JetBrains extensions for inline completion and interactive chat.
Sourcegraph integrated it into Cody to trim autocomplete latency, while LangChain and LlamaIndex hooked into its instruct model for tool use and agent workflows.
Licensing and where to run it
Codestral weights are available for download on Hugging Face.
Keep an eye on the fine print: it is released under the Mistral AI Non-Production License. You can freely pull the weights for research, testing, and evaluation, but deploying it in a commercial product requires a separate commercial license.
If you prefer cloud access, Mistral opened a dedicated endpoint at codestral.mistral.ai, tailored for IDE integrations and personal keys with an 8-week free beta waitlist. Standard pay-per-token API access is also live on api.mistral.ai.
Why it matters
At 22B parameters, Codestral hits an ideal middle ground. It delivers the low latency required for real-time editor autocomplete without sacrificing the smarts needed for complex syntax across dozens of languages.
The non-production license means companies cannot quietly ship it inside revenue-generating software for free. But for individual builders looking for a fast, responsive copilot alternative inside VS Code or JetBrains, jumping into the 8-week free beta is a no-brainer.
Sources (1)
- mistral.ai(opens in a new tab)Originalmistral.ai