Mistral: Codestral 2508 (batch)
About this model
Mistral Codestral 2508 (batch) is the asynchronous batch variant of Codestral 25.08, Mistral’s July 2025 update to its proprietary code-completion family. The model targets low-latency, high-frequency developer workflows such as fill-in-the-middle (FIM) completion, inline edits, code correction, and test scaffolding across 80+ programming languages. It is exposed through Mistral’s hosted API (including the dedicated FIM endpoint), with batch pricing for throughput-oriented workloads that do not require real-time responses.
Relative to earlier Codestral releases, Mistral reports stronger production signals including higher accepted completions, more retained suggestions, and fewer runaway generations, alongside gains on academic FIM benchmarks and chat-mode improvements (+5% on IFEval v8 and +5% average on MultiplE). The stack integrates with Mistral Code for JetBrains and VS Code, and supports structured outputs, function calling, document Q&A, and prefix conditioning within a 128K-token context window.
Codestral 2508 is closed-weight and API-only (no public Hugging Face weights). The batch SKU uses the same model capabilities as the standard endpoint with discounted async inference, making it suited to bulk code transformation, repository-scale suggestion jobs, and offline evaluation pipelines while keeping data on Mistral’s enterprise-grade hosted infrastructure.
Benchmark Scores
Technical Specs
- Architecture: Transformer
- Context Window: 128,000 tokens
- Input Modalities: text
Hardware Requirements
- Compute: API only
Pricing
| Input | Output | Currency |
|---|---|---|
| 0.15 / 1M tokens | 0.45 / 1M tokens | USD |