Mistral releases Medium 3.5, a 128B open-weight coding model, and launches cloud remote agents in Vibe

Mistral AI on May 22, 2026 released Mistral Medium 3.5, a 128B open-weight dense model designed for coding, reasoning, and instruction-following in a single set of weights, alongside a new remote-agents capability in Vibe that moves agentic coding tasks from the local machine to the cloud.

What's new

Medium 3.5 is a 128B dense model with a 256k context window, "handling instruction-following, reasoning, and coding in a single set of weights." It runs self-hosted on as few as four GPUs, making enterprise on-premise deployment practical. Open weights are published on Hugging Face under a modified MIT license.

Benchmark performance:

• 77.6% on SWE-Bench Verified, ahead of Devstral 2 and comparable to much larger closed-source models
• 91.4 on τ³-Telecom

API pricing: $1.5 per million input tokens and $7.5 per million output tokens.

The simultaneous launch of remote agents in Vibe is the product-level headline. Until now, Vibe coding agents ran locally. The update moves execution to the cloud: "Coding agents have mostly lived on your laptop. Today we're moving them to the cloud, where they run on their own, in parallel, and notify you when they're done."

The remote agent capability pairs with Medium 3.5 as the default model in Vibe, replacing the previous Devstral 2 baseline.

Mistral also ships Studio MCP Connectors alongside this release — built-in and custom MCP integrations for enterprise data pipelines — giving teams structured access to private data sources without custom glue code.

Context

Medium 3.5 follows Devstral 2 as Mistral's primary coding model. Rather than a mixture-of-experts (MoE) architecture — which Mixtral and several prior Mistral models used — Medium 3.5 is a dense 128B model. Mistral's positioning is that a well-tuned dense model at this scale can handle the full instruction/reasoning/coding surface in a single checkpoint, avoiding the routing overhead and partial-expert activation of MoE designs.

Le Chat's Work mode also expands with this release, incorporating the remote-agents infrastructure for long-horizon productivity tasks.

Why it matters

77.6% on SWE-Bench Verified is competitive with closed-source frontier models at a fraction of the API cost ($1.5/$7.5 versus $5/$25 for Anthropic's Opus tier). For teams that need strong coding agents and can run open weights, Medium 3.5 changes the cost calculus significantly.

The shift to cloud-native remote agents in Vibe is a direct response to the GitHub Copilot and Cursor model, where coding agents run server-side and handle long-horizon tasks asynchronously. Mistral's open-weight approach means teams can deploy the same model self-hosted in a private cloud, retaining data control while matching the cloud-agent experience.

The sub-four-GPU self-hosting threshold is a meaningful practical constraint: Medium 3.5 fits within a standard two-to-four A100 node, the most common GPU configuration for enterprise on-premise AI deployments.