mistral-medium-3.5

Mistral Medium 3.5 was introduced by Mistral AI as a frontier-class multimodal model aimed squarely at agentic workflows and coding assistance, with image understanding sitting alongside text as a core capability rather than an afterthought. Mistral positions it as the successor to several earlier systems in their lineup, replacing the prior Medium 3.1 model and the Magistral reasoning model inside Le Chat, and taking over from Devstral 2 as the engine behind their Vibe coding agent. The practical effect is a single model that the developer can lean on for interactive chat, code generation and repair, reasoning-heavy prompts, and tool-driven automation without stitching together separate specialists for each task.

Under the hood, the model is described as Mistral AI's first flagship merged model, consolidating instruction-following, reasoning, and coding abilities into one set of weights. It is a dense 128-billion-parameter architecture paired with an extended context window, which makes it well suited to long-document reasoning, multi-step agent loops, and repository-scale coding tasks where keeping large amounts of code, tool traces, or retrieved context in view matters. A separately distributed NVFP4 quantization from NVIDIA shows that the weights can be efficiently compressed for inference deployments while preserving the unified behavior, reinforcing that the model is intended for both cloud serving and more resource-constrained production environments.