Codestral 25.01
by Mistral AI·Europe·Released
Mistral's coding-specialist model — 256K context, fast autocomplete, 80+ languages.
About this model
Codestral 25.01 (January 2025) is Mistral's coding specialist — a 22B-parameter model with a 256K context window and a focus on real-time autocomplete latency rather than chat-style coding. Codestral powers the inline completions in major IDE plugins (Cursor, Continue, Tabnine integrations, JetBrains AI).
Compared to general-purpose models, Codestral is optimised for the fill-in-the-middle (FIM) format that autocomplete engines use, with much lower latency at the cost of being less suited to standalone chat workflows.
Strengths
- •Best-in-class latency for IDE autocomplete (FIM format)
- •256K context — full files / multi-file context in single requests
- •Supports 80+ programming languages
- •Drop-in replacement for older Codestral in major IDE plugins
Limitations
- •Not designed for chat-style coding — use Mistral Large for conversations
- •Weights available only for non-commercial evaluation (Mistral NPL)
- •Beaten by Sonnet 4 / GPT-4.1 on long-horizon agentic coding
When to use it
- →IDE autocomplete (Cursor inline, Continue, JetBrains AI)
- →Fill-in-the-middle code completion at scale
- →Inline refactoring suggestions
- →Multi-file code completion within a 256K window
Architecture & training
22B-parameter dense transformer trained on a code-heavy corpus across 80+ languages. The fill-in-the-middle training data is explicitly weighted to optimise for the autocomplete use case rather than chat. Mistral's technical post emphasises latency-per-completion as the primary optimisation target.
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| MBPP | 80.2 | |
| HumanEval | 86.6 | |
| RepoBench | 38.0 |