Magistral Small 1.1: Enhanced Reasoning Model

Magistral-Small-2507 is a 24 billion parameter language model from Mistral AI, an evolution of Magistral Small 3.1. It incorporates advanced reasoning capabilities, undergoing Supervised Fine-Tuning (SFT) from Magistral Medium traces and Reinforcement Learning (RL). This model is designed to be efficient, capable of local deployment on a single RTX 4090 or a 32GB RAM MacBook after quantization.

Key Capabilities

Advanced Reasoning: Excels at generating long, detailed chains of reasoning traces to arrive at an answer, utilizing [THINK] and [/THINK] special tokens for structured thought processes.
Multilingual Support: Proficient in dozens of languages, including English, French, German, Japanese, Chinese, and Arabic.
Flexible Context Window: Features a 128k context window, with optimal performance recommended up to 40k tokens.
Improved Behavior: Offers better tone, enhanced LaTeX and Markdown formatting, and reduced likelihood of infinite generation loops compared to previous versions.
Apache 2.0 License: Provides an open license for both commercial and non-commercial use and modification.

Performance Highlights

Magistral Small 1.1 demonstrates strong performance across various benchmarks, including AIME24 pass@1 (70.52%), AIME25 pass@1 (62.03%), GPQA Diamond (65.78%), and Livecodebench (v5) (59.17%).

Recommended Usage

For optimal results, users are advised to employ a specific system prompt that encourages an inner monologue and structured reasoning. The model is compatible with vLLM (recommended for inference) and transformers, with community-supported quantized versions available for llama.cpp and lmstudio.