Mistral Medium 3.5 represents Mistral's first "fully merged" hybrid model, combining capabilities from three previously distinct lineages — the Instruct family, the Reasoning line (formerly known as Magistral), and Devstral for coding — into a single unified system. Built on a Dense architecture with 128 billion parameters, the model is designed to handle instruction-following, structured reasoning, and code generation tasks within one inference path, avoiding the overhead of routing between separate specialized models. Mistral positions it as a frontier-class multimodal release optimized for agentic and coding use cases, and distributes it under a Modified MIT license.

In practical terms, the model accepts both text and image inputs while producing text output, making it suitable for workflows that combine visual context with code or tool-based reasoning. It supports a 256k token context window and offers native function calling with JSON-structured output, enabling reliable integration into agent pipelines that require tool dexterity and programmatic responses. Users can switch between an instant-reply mode and a configurable reasoning mode that allows test-time compute scaling for harder problems. The architecture's broad multilingual coverage and "strong tool dexterity" make it a practical fit for agentic research tasks, VIP-level vibe coding scaffolds, and enterprise assistants that need both vision understanding and reliable structured tool use in one model.