MistralAI released a new Magistral Small 2509 (huggingface.co)

MistralAI has released Magistral Small 2509, a 24B-parameter reasoning-focused model built on Mistral Small 3.2 (2506) and further refined with supervised fine-tuning using Magistral Medium traces plus an additional reinforcement-learning stage. The release brings significant capability upgrades: a vision encoder for multimodal reasoning, a 128k context window (recommended effective range ≤40k for best performance), improved tone/persona, reduced infinite-generation risk, and explicit "[THINK]...[/THINK]" tokens and a system-level reasoning prompt to expose and structure long chains of internal reasoning. The model is Apache 2.0–licensed and documented in the Magistral paper and blog post. Technically, Magistral Small 1.2 (the shipped variant) shows large benchmark gains over prior Small 1.1/1.0—e.g., AIME24 pass@1 jumps to ~86.1% from ~70.5%—and narrows the gap with Medium variants. It’s optimized for local deployment (quantized GGUF builds can run on a single RTX 4090 or a 32GB MacBook) and is supported across vLLM (recommended), transformers, llama.cpp, and community GGUF repos. Multilingual coverage spans dozens of languages, and the model’s reasoning-trace tooling plus vLLM streaming support makes it attractive for applications requiring auditable chains of thought, multimodal QA, and long-context workflows.