mistral-small-2506 is a 24-billion-parameter language model developed by Mistral AI, released in June 2025 as part of the Mistral Small 3.2 family. Built as an upgrade to the 3.1 series, this model maintains a compact footprint designed for efficient local deployment and high-throughput enterprise applications while introducing native multimodal capabilities for processing both text and visual inputs.
Technical Capabilities
Architecturally, the model is a decoder-only Transformer featuring a 128,000-token context window, allowing for the processing of extensive documents and multi-turn conversations. While the theoretical limit is high, performance is optimized for best results within the first 40,000 tokens. The 2506 update specifically improved instruction-following accuracy to approximately 84.78% and significantly reduced repetition errors compared to previous versions.
Multimodal and Reasoning Features
The model incorporates image-text-to-text capabilities, enabling tasks such as document understanding, chart analysis, and visual question answering. It was released alongside a reasoning-optimized sibling, Magistral Small, though the standard 2506 model provides a balance of general-purpose utility, tool-calling robustness, and multilingual support across more than 20 languages. The model is released under an Apache 2.0 license, supporting both commercial and non-commercial use.