Mistral has released Devstral 2, a new generation of open-weight coding models designed for agentic software engineering workflows, alongside a companion command-line tool, Mistral Vibe. The release targets organizations seeking production-grade code automation with greater control over cost, deployment, and governance than closed-source alternatives.
The Devstral 2 family consists of two models: Devstral 2, a 123B-parameter dense transformer released under a modified MIT license, and Devstral Small 2, a 24B-parameter model released under Apache 2.0. Both models support a 256K token context window and are positioned for use in large codebases where long-range context, dependency tracking, and multi-file reasoning are operational requirements rather than optional features.
On SWE-bench Verified, a benchmark designed to evaluate real-world software engineering tasks, Devstral 2 achieves 72.2 percent accuracy, while Devstral Small 2 reaches 68.0 percent. These results place both models competitively among much larger proprietary and open models, despite having significantly fewer parameters. The smaller footprint lowers infrastructure requirements and expands deployment options, including on-premise environments and developer workstations, which is particularly relevant for enterprises operating under data residency, security, or cost constraints.
Devstral 2 is optimized for data center deployments and requires a minimum of four H100-class GPUs, while Devstral Small 2 is designed to run on a single GPU and can operate on consumer-grade hardware or CPU-only systems. This range allows organizations to standardize on a single model family across development, testing, and production, while adjusting deployment profiles based on workload criticality and budget.
From a workflow perspective, Devstral 2 is designed to function as a code agent rather than a passive assistant. It can explore entire repositories, reason across multiple files, track framework-level dependencies, and iteratively correct failures. These capabilities align with enterprise use cases such as legacy system modernization, large-scale refactoring, and automated bug resolution. Both models support fine-tuning, enabling organizations to optimize for specific programming languages, internal frameworks, or architectural patterns.
Mistral has also introduced Mistral Vibe, an open-source command-line interface built specifically for Devstral. Vibe operates directly in the terminal or via IDE integrations using the Agent Communication Protocol, enabling natural-language driven code exploration, modification, and execution. The tool maintains project-level awareness by scanning file structures and version control state, supports multi-file orchestration, and exposes controls for tool permissions and automation behavior. For teams experimenting with agentic development, Vibe provides a reference implementation that can be adapted or embedded into existing toolchains.
In comparative evaluations conducted with independent human annotators, Devstral 2 outperformed DeepSeek V3.2 in head-to-head tasks but remained less preferred than Claude Sonnet 4.5, highlighting both the progress and remaining gap between open-weight and leading proprietary models. For many enterprises, however, the trade-off may be acceptable given the benefits of open licensing, deployment flexibility, and predictable costs.
Devstral 2 is currently available for free via Mistral’s API, with usage-based pricing planned after the introductory period. Devstral Small 2 is offered at lower token costs and is also intended for local and private deployments. The models are integrated with existing open agent frameworks, including Kilo Code and Cline, and Vibe is available as an extension for the Zed editor.
Taken together, the Devstral 2 release positions Mistral as a supplier of open, production-oriented coding models aimed at organizations that want to operationalize AI-driven software engineering without fully outsourcing control to proprietary platforms.