Mistral Medium 3.5
2026-05-22 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI introduces Mistral Medium 3.5, its first merged flagship model: 128 billion dense parameters, 256k context, instruction, reasoning, coding and vision, in open weights under a modified MIT license.
What changes
- Merges instruction, reasoning and coding into a single set of weights (replaces Devstral 2 in Vibe CLI).
- Reasoning configurable per request (reasoning_effort parameter).
- Vision encoder trained from scratch for variable image sizes and formats.
- Default model of Le Chat and Mistral Vibe, engine of Vibe's remote agents.
- Self-hosting possible on four GPUs.
### Specifications - Parameters: 128B dense (no mixture of experts). Context: 256k tokens. - API identifier: mistral-medium-3-5. Version 26.04. - License: modified MIT (open weights); according to the press, companies exceeding about 20 M USD in monthly consolidated revenue need a commercial license (threshold not verified against the license itself). - Availability: Mistral API, Le Chat (Pro, Team, Enterprise), Vibe, Hugging Face, NVIDIA NIM. - Knowledge cutoff: not disclosed. ### Pricing | Item | USD per million tokens | |---|---| | Input | 1.50 | | Output | 7.50 | ### Benchmarks | Benchmark | Score | |---|---| | SWE-Bench Verified | 77.6% | | τ³-Telecom | 91.4 | Scores published by the vendor. It is compared with Devstral 2 and Qwen3.5 397B A17B. ### Key takeaways A single open model for chat, coding and long-running agents, with a moderate API cost, simplifies the catalog of models to manage.
Links
Tags : LLM, Mistral, Medium, Open Source, Agents, Code