Mistral Medium 3.5 is a 128B dense model merging coding, reasoning, and instruction-following in one set of weights. 256k context, configurable reasoning effort. Open weights on HuggingFace for engineers and teams running self-hosted inference.