Atlas

Models

← All models

Codestral 25.08

Mistral AI · 2025-07-30

Codestral 25.08 (`codestral-2508`) is Mistral AI's July 2025 update to its specialized code-completion model, built for high-precision fill-in-the-middle (FIM) completion and code generation in latency-sensitive production environments. Mistral reported over prior versions a 30% increase in accepted completions, 10% more retained code, 50% fewer runaway generations, and gains in FIM benchmarks and chat mode; the current model card labels its context as 128K tokens.

Benchmark scores

BenchmarkScore
ALE-Bench341
BuseyBench SVG2.5
Capability129.9
Design Arena Overall Code Categories1037
EDIT-Bench Complete44.8
EDIT-Bench Complete Easy70.8
EDIT-Bench Complete Hard13.5
EDIT-Bench Core43.5
EDIT-Bench Core Easy64.5
EDIT-Bench Core Hard15.2
LisanBench Difficulty-Weighted13.2
LisanBench Path Length44.0
ObviousBench41.0
Ruby LLM Benchmarks — Program Fixer67.1
Ruby LLM Program Fixer — Calendar85.3
Ruby LLM Program Fixer — Parking Garage60.6
Ruby LLM Program Fixer — School Library47.2
Ruby LLM Program Fixer — Vending Machine75.4
Loading Atlas data…