Magistral Small 1.2 Overview

Magistral Small 1.2 is a 24 billion parameter multimodal language model from Mistral AI, based on Mistral Small 3.2 (2506). It has been fine-tuned with added reasoning capabilities through Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), making it an efficient model for complex analytical tasks. A key update in this version is the integration of a vision encoder, allowing it to process and reason based on multimodal inputs, including images.

Key Capabilities

Advanced Reasoning: Capable of generating long chains of reasoning traces before providing an answer, encapsulated by [THINK] and [/THINK] special tokens.
Multimodality: Features a vision encoder for analyzing images and reasoning from visual content, alongside text.
Multilingual Support: Supports dozens of languages, including English, French, German, Japanese, Chinese, and Arabic.
Extended Context Window: Offers a 128k context window, with good performance up to 40k tokens.
Improved Performance: Demonstrates significantly better performance compared to Magistral Small 1.1 across various benchmarks like AIME24, AIME25, GPQA Diamond, and Livecodebench.
Enhanced Output Quality: Provides better LaTeX and Markdown formatting, shorter answers for easy prompts, and reduced likelihood of infinite generation loops.

Good For

Applications requiring robust reasoning and problem-solving.
Multimodal tasks involving both text and image analysis.
Deployments on resource-constrained hardware like a single RTX 4090 or a 32GB RAM MacBook, once quantized.
Developers looking for an open-licensed model (Apache 2.0) for commercial and non-commercial use.