Mistral AI Releases Magistral Reasoning Models: 24B Open-Source and Enterprise Versions Score 70.7% and 73.6% on AIME202
Mistral AI has released Magistral, its first reasoning model line, in two versions: Magistral Small (24B parameters, Apache 2.0) and Magistral Medium (enterprise). Magistral Medium scored 73.6% on AIME2024 (90% with majority voting at 64 samples), while the open-source Small version achieved 70.7% (83.3% with voting).
Mistral AI Releases Magistral Reasoning Models: 24B Open-Source and Enterprise Versions Score 70.7% and 73.6% on AIME2024
Mistral AI has released Magistral, its first reasoning model line, in two versions: Magistral Small (24B parameters, Apache 2.0) and Magistral Medium (enterprise). According to Mistral AI, Magistral Medium scored 73.6% on AIME2024 (90% with majority voting at 64 samples), while the open-source Small version achieved 70.7% (83.3% with voting).
Technical Specifications
Magistral Small:
- 24 billion parameters
- Apache 2.0 license (open-source)
- Available on Hugging Face for self-deployment
- AIME2024: 70.7% (single attempt), 83.3% (majority voting @64)
Magistral Medium:
- Parameter count not disclosed
- Enterprise-only version
- AIME2024: 73.6% (single attempt), 90% (majority voting @64)
- Available via La Plateforme API, Le Chat, Amazon SageMaker
- Coming soon to IBM WatsonX, Azure AI, Google Cloud Marketplace
Core Capabilities
Both models feature chain-of-thought reasoning that operates natively across multiple languages and alphabets, including English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. Mistral AI emphasizes the models are designed for multi-step logic with transparent, traceable thought processes.
The company claims Magistral Medium achieves up to 10x faster token throughput than most competitors when using "Flash Answers" mode in Le Chat, though specific token-per-second figures were not provided.
Target Applications
Mistral AI positions Magistral for:
- Legal research and compliance (with auditable reasoning chains)
- Financial forecasting and risk modeling
- Software development and system architecture
- Strategic planning and operational optimization
- Creative writing and content generation
The models are designed for use cases requiring "longer thought processing and better accuracy than with non-reasoning LLMs," according to the company.
Pricing and Access
Pricing for Magistral Medium API access was not disclosed. Magistral Small is free to download and deploy under Apache 2.0 license. Enterprise customers can request on-premises deployments through Mistral's sales team.
Research Publication
Mistral AI released an accompanying research paper covering training infrastructure, reinforcement learning algorithms, and evaluations. The company stated it aims to "iterate the model quickly" with constant improvements expected.
What This Means
Mistral's entry into reasoning models with an open-source 24B parameter option creates a new benchmark for accessible chain-of-thought models. The AIME2024 scores place both versions competitively within the current reasoning model landscape, though direct comparisons require knowing context windows and pricing structures still undisclosed for the Medium version. The multilingual chain-of-thought capability addresses a genuine gap in existing reasoning models, which typically perform best in English. However, the 10x speed claim lacks the specific comparative data needed for verification.
Related Articles
miHoYo-linked team open-sources GameScaling: eight checkpoints, 800M to 27B parameters, play games via keyboard and mous
A research team involving miHoYo's Honkai AI researchers and university collaborators has open-sourced GameScaling, a family of eight checkpoints from 800 million to 27 billion parameters. The models generate keyboard and mouse actions from visual observations and were built by continued pretraining of Qwen3.5 vision-language models.
MOSI and OpenMOSS open-source MOSS-Image, a 5B text-to-image model for dense text and JSON layout control
MOSI Intelligence and OpenMOSS have open-sourced MOSS-Image, a 5B-parameter text-to-image model focused on dense text and JSON layout control. According to Pandaily, it quickly topped Hugging Face's text-to-image trending list.
Gradium open-sources Phonon 31m, a 31M-parameter on-device TTS model under Apache 2.0
Gradium has open-sourced Phonon, a family of on-device text-to-speech models. The English Phonon 31m has 31M parameters, ships under Apache 2.0, fits in 64.5 MB, and starts speaking in about 7 ms on a laptop CPU.
KRAFTON releases Raon-1 mini preview, a 28.8B-parameter MoE model with 4.3B active and 262,144-token context
KRAFTON AI has released Raon-1 mini preview, a Mixture-of-Experts language model trained from scratch with 28.8B total and 4.3B active parameters. It is Apache 2.0 licensed, supports 262,144 tokens of context, and ships with a speculative decoding draft model that KRAFTON says reaches up to 2.18× higher output throughput on a single NVIDIA B300.
Comments
Loading...