French startup Mistral AI introduced a new reasoning model designed for domain-specific, transparent and multilingual reasoning.
The AI vendor launched two versions of Magistral: Small and Medium.
Magistral Small, the open source version, is a 24 billion-parameter model. Magistral Medium is the enterprise version. It is available on Amazon SageMaker and will soon be available on IBM Watsonx, Microsoft Azure AI and Google Cloud Marketplace.
The large language model works across global languages and alphabets and is suited for structured calculations and programmatic logic, Mistral said. The model can reason in French, Spanish, German, Italian, Arabic, Russian, English and Simplified Chinese.
Magistral represents a trend among frontier model makers to focus on test-time reasoning, according to Bradley Shimmin, an analyst at The Futurum Group. Test-time reasoning is a technique that enables a model to engage in a deliberate thinking process when answering questions, instead of just regurgitating information.
Magistral's introduction comes after Chinese AI vendor DeepSeek introduced its reasoning model in January and rival OpenAI launched two new reasoning models in April.
"These companies are exploring how to improve that ability," Shimmin said. "It is true that a reasoning model is going to be more accurate ... than a non-test-time reasoning model."
With Magistral, Mistral is focusing on performance, while still trying to maintain its stance as an open model provider with the release of Magistral Small as an open weight model under the Apache 2.0 license, he said.
Traceable reasoning and a challenge
While the Magistral model is comparable to other reasoning models such as DeepSeek-R1, it stands out in that it can provide a traceable reasoning or thought process.
"Instead of showing a polished chain of thought, traceable reasoning requires showing the complete flow of confidence scores, alternative hypotheses and even the most inelegant error-correction steps," said Torsten Volk, an analyst at Enterprise Strategy Group, now part of Omdia.
Auditors only care about completeness, not elegance, when verifying compliance.
Torsten VolkAnalyst, Enterprise Strategy Group
According to Mistral, the importance of traceable reasoning is that professionals in the legal, finance and healthcare industries and government can meet compliance requirements.
"Auditors only care about completeness, not elegance, when verifying compliance," Volk said.
Therefore, he added, if Mistral decides to become a compliance leader in the LLM market, it could be a big deal because many projects are shelved before they start due to legal problems.
"Supporting an open source community that helps ensure the completely open character of Magistral could open up many of these previously taboo use cases," Volk said.
Magistral Medium is also able to provide faster reasoning with Flash Answers in Mistral's Le Chat chatbot. This allows for real-time reasoning and user feedback, according to the vendor.
Volk said the real-time reasoning capability could be helpful in various areas, including customer service, when a representative needs to respond to a customer in real time.
While showing promise, Mistral's biggest challenge is rising above the "consistent onslaught of marketing hype that their competitors are willing to invest in," Shimmin said. He added that while the vendor has some support from Microsoft, it does not mean that enterprises will readily adopt Mistral models.
Other than a reasoning model, Mistral on Wednesday launched Mistral Compute, an AI infrastructure platform powered by Nvidia's latest Grace Blackwell chip.
Esther Shittu is an Informa TechTarget news writer and podcast host covering AI software and systems.
