Magistral

2025-06-10 · Publisher : Mistral AI · Domain : Artificial Intelligence

Mistral AI introduces Magistral, its first family of reasoning models: Magistral Small (24B, Apache 2.0) and Magistral Medium (enterprise version), with traceable, multilingual reasoning.

Mistral AI introduces Magistral, its first family of reasoning models: Magistral Small (24B, Apache 2.0) and Magistral Medium (enterprise version), with traceable, multilingual reasoning. ### What changes - Mistral's first reasoning offering, in two variants: open Small, proprietary Medium. - Traceable chain of thought in the user's language (English, French, Spanish, German, Italian, Arabic, Russian, Simplified Chinese). - Think mode and Flash Answers in Le Chat, with a throughput announced at up to 10 times higher. - Targets: legal, finance, healthcare, software engineering, planning. ### Specifications - Magistral Small: 24 billion parameters, Apache 2.0 license, weights on Hugging Face (Magistral-Small-2506). - Context: 128k for Small, with possible degradation beyond 40k (HF card); documentation: 40k for Magistral Medium 1.0. - Magistral Medium: proprietary, in preview at launch; Amazon SageMaker, then IBM watsonx, Azure AI and Google Cloud announced. - API identifiers: `magistral-medium-2506`, `magistral-small-2506`. ### Pricing | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Magistral Medium | 2 | 5 | | Magistral Small | 0.5 | 1.5 | Prices from the pricing page archived on June 20, 2025; not written in the post. ### Benchmarks Scores published by the vendor: | Benchmark | Medium | Small | |---|---|---| | AIME 2024 (pass@1) | 73.6% | 70.7% | | AIME 2024 (majority vote @64) | 90% | 83.3% | | AIME 2025 (pass@1) | 64.95% | 62.76% | | GPQA Diamond | 70.83% | 68.18% | | LiveCodeBench (v5) | 59.36% | 55.84% | ### Key takeaways A reasoning model whose logic is auditable, with an open version that can be deployed in-house and a managed version for production.

Tags : LLM, Mistral, Magistral, Raisonnement, Open Source