The larger model, Devstral 2, has 123 billion parameters. Its benchmarks match those of DeepSeek V3.2 and Kimi K2, but it's much smaller than its competitors. It runs on four Nvidia H100 chips. It's licensed under the MIT license.

Devstral 2 is temporarily free in the API. After the promotional period, the price will be $0.40/$2 per million tokens.

The smaller model, Devstral Small, has 24 billion parameters. It runs on a powerful GeForce RTX. It's available under the Apache 2.0 license. The API price is $0.10/$0.30 per million tokens.

Vibe CLI, an AI agent for terminal-based development, was released alongside these models.