Introducing: Devstral 2 and Mistral Vibe CLI.
The coding agent and its model family
Mistral introduced Devstral 2 and Mistral Vibe CLI in its December 2025 release article. Devstral 2 is the 123-billion-parameter model, while Devstral Small 2 is the 24-billion-parameter version. The article gives the larger model a modified MIT licence and the smaller one Apache 2.0.
Vibe CLI is the associated command-line coding assistant, not another name for the model weights. Mistral's current plans page still lists a terminal-native Vibe CLI and separates it from web, mobile and IDE interfaces. The launch article should not be read as proof that Devstral 2 is Mistral's newest served model or that its introductory free API offer remains active.
Work across a repository
The release article describes an interactive chat interface with tools for manipulating files, searching code, working with version control and executing commands. Project-aware context includes the file structure and Git status. It also describes file references through autocomplete, shell-command shortcuts, persistent history and programmatic use for scripting.
Tool permissions are configurable, including an auto-approval setting. For an initial evaluation, keep approvals deliberate and choose a bounded repository task with a clear acceptance test. Ask for a small change, inspect the diff, and run the relevant checks before trusting a larger refactor. This is suggested workflow guidance, not evidence of a measured improvement in development speed.
Local models and current access
The release article describes configuration of local models and providers through a configuration file. It distinguishes the larger Devstral deployment, aimed at data-centre GPUs, from Devstral Small 2, which supports single-GPU and CPU-only configurations. Availability on consumer hardware is not a promise of a particular speed or memory footprint on your machine.
Current access conditions belong to the account and model route you choose. The plans page describes limited free coding sessions and higher paid-plan access, subject to usage conditions. Check current billing and model availability instead of copying token prices from the launch announcement.
For a useful trial, compare the agent's edits with your repository conventions, review commands before they run, and confirm that tests actually execute. Official product descriptions establish the intended capabilities; this entry has not tested autonomous changes or benchmark performance.