Mistral has done some spring cleaning. Instead of maintaining three specialized models - Devstral 2 for coding, Magistral for reasoning, Medium 3.1 for general instructions - there’s now one: Mistral Medium 3.5. A dense model with 128 billion parameters and a 256,000-token context window.
Why a Dense Model?
The decision against Mixture-of-Experts is deliberately conservative. Competitors like DeepSeek and Qwen have been moving their flagship models toward MoE architectures that deliver cheaper inference at similar quality. Mistral’s own Large 3 uses 675 billion parameters but only activates 41 billion per token.
Medium 3.5 activates all 128 billion parameters for every single token. That makes inference more expensive, but the model is simpler to deploy and maintain. Mistral says four GPUs are enough for self-hosting - though ‘four GPUs’ in practice still means a data center.
Reasoning as a Toggle
Instead of a separate reasoning model, there’s now a reasoning_effort parameter. Simple queries get fast answers, complex agent tasks get the full reasoning mode. On top of that, Mistral retrained the vision encoder from scratch to handle variable image sizes and aspect ratios.
In Mistral’s own benchmarks, Medium 3.5 scores 77.6% on SWE-Bench Verified and 91.4% on their telecom industry benchmark. In banking scenarios, though, Claude is significantly ahead.
Modified MIT Instead of Apache 2.0
The license change is worth noting. Previous Mistral models like Large 3 and Small 4 shipped under Apache 2.0. Medium 3.5 comes with a ‘Modified MIT License’ that allows commercial and non-commercial use but carves out exceptions for high-revenue companies. Mistral is moving toward the kind of licensing models that Meta uses with Llama.
Vibe Gets Cloud Agents
The arguably more interesting news is the update to Mistral’s coding tool Vibe. It’s getting asynchronous cloud agents that work independently in isolated sandboxes and can open pull requests when they’re done. Integrations with GitHub, Linear, Jira, Sentry, Slack, and Teams are built in. The concept is familiar from Claude Code and OpenAI Codex - but Mistral is the first European company to join that league.
Le Chat Becomes a Work Assistant
Le Chat, Mistral’s consumer app, gets a ‘Work Mode’ for multi-step tasks. Connectors to email, calendars, and document systems are enabled by default. Sensitive actions like sending messages require explicit confirmation. It sounds a lot like Anthropic’s Cowork approach, just with European infrastructure.
Through the API, Medium 3.5 costs $1.50 per million input tokens and $7.50 per million output tokens.
Sources: Mistral AI: Vibe Remote Agents and Medium 3.5, The Decoder: Mistral Medium 3.5, Hugging Face: Mistral Medium 3.5 Model Card