Magistral

lastmod 2025-06-11created 2025-06-111 edits

Mistral AI's first reasoning model

About

Magistral is Mistral AI's first reasoning model, specialized in domain-specific, transparent, and multilingual reasoning
Released in two versions: the open-source Magistral Small (24B parameters) and the enterprise version Magistral Medium
Chain of Thought-based multilingual reasoning, transparently providing the step-by-step logic in the user's language
On AIME2024, Magistral Medium scored 73.6% (up to 90%) and Small 70.7% (up to 83.3%)
Supports precise step-by-step reasoning and 10x faster response speed suited to various languages and industries, including regulated industries such as law, finance, and healthcare, data engineering, software development, and creative content

Magistral is a reasoning model focused on real-world problem solving and feedback-based improvement
Dual release: Magistral Small, a 24B-parameter open-source version, and Magistral Medium, a more powerful enterprise version
Performance metrics:
- Magistral Medium: 73.6% on AIME2024, 90% with majority voting
- Magistral Small: 70.7% and 83.3%, respectively
Applies Chain of Thought reasoning across global languages and scripts, enabling native-level thinking
Suited to a range of tasks such as structured calculations, programming logic, decision trees, and rule-based systems
Le Chat's Think mode and Flash Answers features deliver 10x faster responses than competitors
The official paper includes evaluations covering the algorithms, training infrastructure, reinforcement learning techniques, and training insights

Transparent reasoning process:
- Magistral is optimized for multi-step logic, so users can check and trace the reasoning process in their own language
- Stronger interpretability and verifiability than typical models
- Aims for continuous model updates and rapid improvement
Multilingual reasoning: high accuracy and preserved logic in English, French, Spanish, German, Italian, Arabic, Russian, Chinese, and more
Response speed:
- Magistral Medium, via Le Chat's Flash Answers feature, supports real-time reasoning and feedback with 10x higher token throughput than competitors
- Demonstrated superior speed compared with major competing models such as ChatGPT

Magistral Small is released under the Apache 2.0 license
Users can directly analyze, modify, and rebuild its architecture and reasoning approach
Previous open-source models have been used in innovative research projects such as ether0 and DeepHermes 3

Magistral is optimized for domains where precise step-by-step reasoning and transparency matter, such as law, finance, software development, and storytelling
Business strategy and operations
- Can perform strategic planning, risk assessment, data-driven decision-making, and computing optimal solutions under complex constraints
Regulated industries and the public sector
- Legal, finance, healthcare, and government professionals can trace logical reasoning paths and ensure auditability
- Helps meet auditability of results and regulatory compliance
Systems, software, and data engineering
- Improved quality of support for programming, project design, backend architecture, and data engineering compared with non-reasoning LLMs
- Effective for multi-step tasks such as working with external tools and APIs
Content creation and communication
- Magistral also shows excellent results in creative writing and storytelling
- Can generate not only coherent text but also unique and quirky ideas

The Small version can be downloaded and self-deployed
The Medium version is available now on Le Chat (web), the API, and Amazon SageMaker
Support on IBM WatsonX, Azure AI, and Google Cloud Marketplace coming soon
Contact separately for enterprise-custom and on-premises deployments

author Zer0 sitemap