Mistral launches Devstral 2, a coding model with 123B parameters requiring at least four H100 GPUs, and a 24B parameter Devstral Small for local deployment
French AI startup Mistral today launched Devstral 2, a new generation of its AI model designed for coding, as the company seeks …
The 123B scale also reprises Mistral’s earlier flagship-model size tier, while its prior workstation and server fine-tuning tools indicate an existing emphasis on deployable customization rather than a single hosted endpoint.
First-order effects
Teams adopting Devstral 2 must have access to at least four H100 GPUs, making its use an infrastructure decision as well as a model-selection decision.
Devstral Small gives users seeking local deployment a smaller alternative within the same coding-focused product line.
Second-order effects
Mistral’s two-model lineup segments coding workloads between GPU-intensive centralized serving and local installations, requiring buyers to weigh hardware availability against where they want inference to run.
The large model’s stated hardware floor makes accelerator access more consequential for organizations that want to operate Mistral’s highest-scale coding model themselves.
Third-order effects
If this portfolio structure persists, coding AI may increasingly be sold as paired deployment tiers: large models for well-provisioned infrastructure and smaller models for local control.
That split could make inference infrastructure a more durable source of differentiation among model providers, even when models target the same developer workflows.
The trend: The release is part of an AI hardware strategy split in which model vendors pair frontier-scale systems with smaller deployable variants to serve uneven access to compute.
Glad to release the new devstral, the best open model for code agents! It comes with vibe CLI, and you can install it easily with uv or pip and let it guide you! uv tool install mistral-vibe pip install mistral-vibe
Our next-generation coding model family Devstral 2 is available in two sizes: Devstral 2 (123B) under a modified MIT license, and Devstral Small (24B) under Apache 2.0. Both SOTA, open-source, free to use, and available now via our API. [image]
Mistral Vibe CLI is an open-source command-line coding assistant powered by Devstral. It explores, modifies, and executes changes across your codebase using natural language. Also under Apache 2.0. Install via: uv tool install mistral-vibe [video]
I've been lucky enough to use it for a couple of weeks, worth a shot even if you're not into super agentic workloads (I'm not). First time I've been super impressed since Claude Code, Vibe and Devstral are really good
Very excited to release two new open-weight models, Devstral 2 (123B) and Devstral Small 2 (24B), along with Mistral Vibe, a CLI built for Devstral that enables end-to-end code automation! [image]
Check out the new Devstral 2 models. 72.2% SWI-Bench for the 123B (~SoTA amongst any OSS model). But the 68.8% with our 24B is particularly crazy progress for laptop-friendly models over the past year.