One of Europe’s leading AI companies, Mistral, has announced the availability of its Magistral reasoning model. It comes in two versions: there is the open-source Magistral Small (24B parameters) and the enterprise-focused Magistral Medium.
Magistral has been designed to better handle domain-specific tasks such as law, finance, and engineering. It shows users the model’s thought processes so that it’s more transparent, and it supports multilingual reasoning to perform better when tasked with answering non-English queries.
Key features and performance
Mistral compared its new models against DeepSeek R1 in various benchmarks. Ones that stood out were its results on AIME2024, which assesses reasoning capabilities. Magistral Medium scored 73.6% (90% with majority voting @64) while Small scored 70.7% (83.3% respectively).
If you’ve used DeepSeek R1, you’ll have seen it thinking before answering. Magistral now does this too and can do it across multiple languages, including English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. Interestingly, there’s also a Flash Answers mode that speeds up the thinking for 10x faster responses compared to competitors.
Real-world applications and availability
Thanks to the expertise Magistral has been trained on, organizations across multiple fields, including legal research, financial forecasting, and regulated industries, may be able to turn to Magistral for more reliable answers. Its traceable reasoning also makes it ideal in regulated fields for compliance and auditability.
Magistral Small, the open-source model, is available under an Apache 2.0 license via Hugging Face. Magistral Medium is accessible through Mistral's Le Chat and via API on La Plateforme. It’ll also be arriving soon on Amazon SageMaker, IBM WatsonX, Azure AI, and Google Cloud Marketplace.
The easiest way for readers to try it out is on Mistral's Le Chat platform; just head to chat.mistral.ai to get started.
Source: Mistral
4 Comments - Add comment