Mistral Medium 3.5 is a flagship dense language model developed by Mistral AI, released on April 29, 2026. It serves as a unified successor to several of the company's specialized models, including Mistral Medium 3.1, Magistral, and Devstral 2. By consolidating instruction-following, complex reasoning, and coding capabilities into a single 128B parameter set, the model provides consistent performance across diverse agentic and conversational workflows without the architectural complexity of Mixture-of-Experts (MoE) systems.\n\nThe model is built on a dense transformer architecture featuring 88 decoder layers, utilizing Grouped Query Attention (GQA) and a 262,144-token context window. It is natively multimodal, incorporating a Pixtral vision encoder trained from scratch to handle variable aspect ratios and image resolutions. This design provides a notable footprint advantage over comparably capable sparse models, allowing for high-performance inference on relatively compact hardware configurations for a model of its weight class.\n\nA defining feature of Mistral Medium 3.5 is its configurable Reasoning Mode. Users can toggle between a standard "fast" mode for extraction and simple chat, and a high-effort reasoning mode that utilizes test-time compute to solve complex mathematical and logical problems. In benchmarks, the model demonstrates high proficiency in software engineering, scoring 77.6% on SWE-Bench Verified and 91.4% on Α-Telecom. It also features robust multilingual support for over 40 languages and strong adherence to system prompts.\n\nFor optimal performance, Mistral recommends specific settings based on the task. When using the high reasoning effort setting, a temperature of 0.7 is advised, while standard chat tasks perform well at temperatures between 0.0 and 0.7. The model is released under a Modified MIT License, providing open-weight access for researchers and developers, subject to specific revenue-based commercial terms for large-scale enterprise users.