Mistral Medium 3.5

Mistral Medium 3.5 is a 128-billion-parameter dense model designed to consolidate several specialized lineages into a single system. According to the Azure Foundry catalog, it merges capabilities from the Instruct family, the Reasoning family previously known as Magistral, and Devstral, so instruction following, multi-step reasoning, and code generation live in one checkpoint rather than requiring model routing. The catalog also describes it as Mistral's first "fully merged" release, and it adds multimodal input for both text and images alongside native function calling with structured JSON output, positioning it as a general-purpose engine for agentic and coding workflows.

The model's practical profile is shaped by a 256k-token context window, broad multilingual coverage spanning more than forty languages, and a Modified MIT license that lowers friction for self-hosting and customization. Early community testing on NVIDIA DGX Spark and GB10 hardware demonstrates that the 128B weights respond well to NVFP4 quantization combined with the EAGLE speculative-decoding method, suggesting efficient single-node inference is realistic for organizations that want to run it locally. With configurable instant-reply and reasoning modes, Mistral Medium 3.5 is a strong fit for teams building tool-using assistants, long-context document analysis, and code-generation agents that need vision input alongside flexible deployment options.