Mistral Small 3.2 Review — 24B Instruction Refinement, 128K Context, Apache 2.0

Mistral Small 3.2 (June 2025) refines Mistral Small 3.1's 24B dense model with dramatically better instruction following — Arena Hard jumps from 19.56% to 43.10%, Wildbench from 55.60% to 65.33% — while halving the repetition bug rate and improving code benchmarks. MMLU flat at 80.5%. Apache 2.0. Ollama: mistral-small3.2. Rating: 4/5.