Magistral-24B: An Enhanced Reasoning Model
The csalab/Magistral-24B is a 24 billion parameter language model, an evolution of Mistral Small 3.1, specifically enhanced for advanced reasoning. It incorporates Supervised Fine-Tuning (SFT) from Magistral Medium traces and Reinforcement Learning (RL) to improve its ability to generate detailed thought processes before arriving at a final answer.
Key Capabilities
- Advanced Reasoning: Excels at tasks requiring long chains of reasoning, drafting internal thoughts (
<think>) before producing a concise summary. - Multilingual Support: Capable of processing and generating text in dozens of languages, including English, French, German, Japanese, Korean, Chinese, and Arabic.
- Deployment Flexibility: Designed to be efficient enough for local deployment, fitting on a single RTX 4090 or a 32GB RAM MacBook when quantized.
- Open Licensing: Released under the Apache 2.0 license, permitting broad 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-24B demonstrates strong performance in reasoning benchmarks, achieving 70.68% on AIME24 pass@1 and 68.18% on GPQA Diamond, positioning it as a capable model for complex problem-solving. It is recommended to use specific sampling parameters (top_p: 0.95, temperature: 0.7, max_tokens: 40960) and a structured chat template for best results, especially leveraging the system prompt for reasoning traces.