米斯特拉尔人工智能发布了马吉斯特拉尔，其首个以推理为重点的语言模型

Released Model: Mistral AI has released Magistral, a new model family for transparent, multi-step reasoning, with open and enterprise versions.
- Designed For: Structured, interpretable reasoning across various fields like law, finance, healthcare, logistics, and software.
- Language Support: Supports multi-step chain-of-thought generation in multiple languages including Arabic, Chinese, French, German, and Spanish.
Benchmark Performance:
- Magistral Medium scored 73.6% on AIME 2024 and 90% with majority voting @64.
- Magistral Small reached 70.7% and 83.3% respectively.
Key Features: Emphasizes clarity in logic and step-by-step traceability, suitable for auditability-required use cases. Promotes speed as a key differentiator with up to 10x faster token throughput in Le Chat compared to standard models.
Early User Feedback: Some users think 10x inference with only 10% improvement and general usability decline is not worth it. API pricing based on token use also doesn't make sense. One user won't use it again unlike Mistral Small 3.1.
Context Length Limitation: Many enterprise-grade models have context limits beyond 100K tokens, but Magistral currently offers 40K tokens. An AI solutions company COO commented that the use cases may be limited.
Training Focus: Trained with deep reasoning, RLHF, and transparency in multi-step logic. The accompanying research paper outlines training methods and insights.
Deployment Options: Magistral Small is available for self-hosted deployment via Hugging Face. Magistral Medium can be accessed in Le Chat, and further rollout is planned to platforms like Azure AI, IBM WatsonX, and Google Cloud Marketplace. Mistral aims for rapid iteration of the Magistral family and early community interest is expected in building upon the open-weight Small model.