Mistral introduces Devstral 2, a new family of models for writing and orchestrating code, and Mistral Vibe, a native CLI that turns your terminal into a software engineering assistant. If you're interested in automating real tasks in projects, this shifts the landscape for developers, small teams, and anyone wanting to run models locally.
What is Devstral 2 and Devstral Small 2
Devstral 2 is a dense model with 123B parameters and a context window of 256K tokens, released under a modified MIT license. Devstral Small 2 is the smaller variant with 24B parameters, under Apache 2.0, designed to run on consumer hardware or on-prem environments.
Why does it matter? Because both are open source, permissive, and optimized for code agents: they can explore codebases, understand dependencies, and make changes that touch multiple files while keeping architectural context.
Devstral Small 2 also supports image inputs, fast inference, and private on-device runs — ideal for prototypes or teams that need privacy and low latency.
Performance, efficiency and comparisons
- Devstral 2 reaches 72.2% on SWE-bench Verified, placing it among the top open-weight code models.
- Devstral Small 2 scores 68.0% on the same metric, competing with models several times larger.
- Mistral claims Devstral is up to 7x more cost-efficient than Claude Sonnet on real-world tasks.
In human evaluations with an independent provider (and orchestration tasks run with Cline), Devstral 2 outperformed DeepSeek R1 with a clear margin. Still, Claude Sonnet 4.5 remains preferred in head-to-head comparisons, so a gap with some closed models persists.
"Devstral 2 is the frontier of open-source code models... it delivers a tool-call success rate comparable to the best closed models." — Cline.
Mistral Vibe CLI: your agent in the terminal
Mistral Vibe CLI is an open source tool under Apache 2.0 that turns your terminal into an assistant capable of exploring, modifying, and executing changes across an entire project using natural language. Some key capabilities:
- Project-aware context: it scans your file structure and Git state to provide relevant context.
- Smart references: reference files with
@, run shell commands with!, and use slash commands for configuration. - Multi-file orchestration: it understands the whole codebase, not just the open file, which can cut PR times in half on complex tasks.
- Persistent history, autocomplete and customizable themes.
Vibe CLI integrates with IDEs via the Agent Communication Protocol, is programmable, lets you configure providers and models via config.toml, and controls tool permissions to fit your workflow.
Practical use cases
- Have a monorepo with microservices? Devstral can explore dependencies and propose coordinated changes.
- Need to fix recurring bugs or modernize legacy code? The model detects faults, attempts fixes, and retries when necessary.
- Are you a hobbyist with a home GPU or even without one? Devstral Small 2 is designed for you: local, fast, and private.
How to get started and costs
Devstral 2 is available for free now via API. After the free period, pricing will be:
- Devstral 2: $0.40 / $2.00 per million tokens (input / output)
- Devstral Small 2: $0.10 / $0.30 per million tokens (input / output)
Mistral also integrates with popular tools and agents like Kilo Code and Cline, and Vibe CLI is available as an extension in Zed to use it from your IDE.
Deployment recommendations
- Devstral 2 is optimized for datacenters and requires at least 4 H100 GPUs to deploy in production.
- Devstral Small 2 runs on a single GPU and works on many NVIDIA configurations and even CPU-only setups.
- For stable performance, Mistral suggests a temperature of 0.2 and following the best practices of
Mistral Vibe CLI.
Final reflection
This matters because it lowers barriers: competitive models, permissive licenses, and a native CLI that brings agents into the daily workflow. If you work with code at scale, the ability to run an agent that understands your repo and makes coherent changes can save weeks of repetitive work.
Ready to try it on a real project and see if it halves your PRs? If you have a home GPU, the Small version is a very practical entry point.