Mistral Medium 3.5 is a dense 128-billion parameter instruction-following model designed to excel in agentic workflows, coding tasks, and complex multi-step reasoning. Its architecture supports both text and image inputs with text output, featuring a custom vision encoder that handles variable image sizes and aspect ratios—making it well-suited for processing real-world documents, screenshots, and visual problem-solving alongside text. The model prioritizes reliable multi-tool calling and long-horizon task completion, with a configurable reasoning effort that lets users adjust how deeply the model thinks through each request. The dense 128B design reflects a deliberate balance between capability and accessibility, as it can be self-hosted on as few as four GPUs, giving organizations a path to deploy powerful reasoning infrastructure without requiring massive data center resources.
While the exact post-training lineage is not detailed in available sources, the model's benchmark profile tells a clear story of specialization. It ranks at the 91st percentile across eight benchmarks, achieves 92% accuracy on coding tasks and 94% on mathematics, and reaches 99.5% accuracy on general knowledge questions—making it particularly accurate among models operating at its speed tier. The 61% long-context reasoning score demonstrates it can maintain coherence across extended contexts, which is critical for agentic applications where tasks span many steps. Exceptional reliability with a 100% success rate and consistently fast response times round out a profile built for production workflows where models need to execute reliably in automated pipelines rather than just excel on isolated benchmarks.
