Hugging Face Magistral-Small-2506: Enhanced Reasoning and Multilingual Capabilities
Magistral-Small-2506 is a 24 billion parameter model developed by mistralai, building on the foundation of Mistral Small 3.1. It incorporates Supervised Fine-Tuning (SFT) from Magistral Medium traces and Reinforcement Learning (RL) to significantly boost its reasoning capabilities, allowing for long chains of reasoning before generating an answer. The model is designed for efficiency, capable of local deployment on a single RTX 4090 or a 32GB RAM MacBook once quantized.
Key Capabilities
- Advanced Reasoning: Specialized in complex reasoning tasks, enabling detailed thought processes before providing solutions.
- Multilingual Support: Comprehensively supports dozens of languages, including major global languages like English, French, German, Japanese, Korean, Chinese, and Arabic.
- Flexible Licensing: Released under an Apache 2.0 License, permitting both commercial and non-commercial use and modification.
- Context Window: Features a 128k context window, with optimal performance recommended up to 40k tokens.
Performance Highlights
Magistral-Small-2506 demonstrates strong performance across various benchmarks, particularly in reasoning and coding tasks. While slightly below Magistral Medium, it achieves competitive scores, such as 70.68% on AIME24 pass@1 and 55.84% on Livecodebench (v5).
Recommended Usage
For optimal results, users are advised to utilize specific sampling parameters (top_p: 0.95, temperature: 0.7, max_tokens: 40960) and to include the default system prompt for reasoning traces. The model is compatible with various inference and fine-tuning frameworks, including vLLM (recommended for inference), llama.cpp, ollama, axolotl, and unsloth.