Mistral Launches Medium 3.5 and Cloud-Based Async Agents

Mistral AI released Medium 3.5 on April 29 — a 128-billion-parameter dense open-weight model unifying chat, reasoning, and code — alongside cloud-based remote agents for async coding. The release positions Mistral as the leading open-weight challenger to frontier closed models.

Key Points:

• Medium 3.5 specs: 128B dense model, 256k context window, MIT-style open license. Handles instruction-following, reasoning, and code in a single set of weights. API: $1.5/M input, $7.5/M output.

• Cloud-based remote agents: coding sessions now run asynchronously in the cloud, spawned from CLI or Le Chat, with notification on completion. Local sessions can be 'teleported' to the cloud mid-run.

• Work Mode in Le Chat (Preview): a powerful agent for complex multi-step research, analysis, and cross-tool actions, with explicit approval required before sensitive operations.

• Self-hosts on as few as four GPUs, making it viable for private infrastructure deployments.

Medium 3.5's open-weight release with a permissive license at $1.5/M tokens is a direct challenge to the closed-model pricing structure. Organizations with privacy requirements or cost constraints now have a frontier-class open alternative.

Why It Matters: The cloud async agent capability mirrors what OpenAI is doing with Codex — but with an open model. For organizations that want agentic AI without vendor lock-in, Mistral's architecture is worth serious evaluation. The 4-GPU self-hosting option opens private deployment for regulated industries.