French AI company Mistral AI has open-sourced Mistral Medium 3.5, a 128 billion parameter model that punches above its weight. The company also updated its coding agent Mistral Vibe to run in the cloud, allowing multiple agents to execute simultaneously and letting users kick off tasks from Le Chat, its chatbot.
https://twitter.com/mistralvibe/status/2049511752379813968
Mistral Medium 3.5 packs 128 billion parameters with a 256,000 token context window. It's built for coding agents, supports dozens of languages including English, Japanese, and French, and can process images alongside text.
The chart below compares Mistral Medium 3.5 against its predecessors Devstral 2 and Devstral 2 Small. The performance jump is clear.
When pitted against rivals like Claude Sonnet 4.5, Claude Sonnet 4.6, Kimi K2.5, GLM-5.1, and Qwen3.5, Mistral Medium 3.5 came out ahead in most agent benchmarks. It also topped massive Chinese models on math tests.
Mistral AI is keen to emphasize that this level of performance comes from a model that can run on just four GPUs. "It delivers excellent real-world performance while being compact enough to run on minimal hardware," the company says.
