A 128B model for coding, reasoning, and long tasks
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.
Mistral Medium 3.5 enables Work mode in Le Chat, an agentic interface that orchestrates cross-tool workflows, web research, document synthesis, and inbox triage. The model calls multiple tools in parallel, maintains session state across turns, and produces structured output for downstream consumption.
Mistral Medium 3.5 includes a vision encoder trained from scratch to handle variable image sizes and aspect ratios, enabling multimodal agentic tasks within a single model without separate vision components.
As a 128B dense model with configurable reasoning effort per request, Mistral Medium 3.5 merges instruction-following, reasoning, and coding into a single set of weights. It can be self-hosted on as few as four GPUs and is available as open weights, enabling custom agent development with full control.
Mistral Medium 3.5 features a 256k context window, enabling it to handle long-horizon tasks and maintain state across multiple tool calls and conversation turns without losing context or task coherence.
Mistral Medium 3.5 powers remote coding agents in Vibe that run asynchronously in the cloud, allowing developers to spawn long-running tasks from CLI or Le Chat and receive notifications when complete. The model scores 77.6% on SWE-Bench Verified and handles module refactors, test generation, dependency upgrades, and bug fixes reliably.