French AI startup Mistral AI has entered the agentic coding market with the launch of its new model, Devstral 2, specifically engineered for advanced software development tasks.
Unveiled today, Devstral 2 is a 123-billion-parameter model equipped with autonomous software engineering capabilities. According to the company, it rivals top-tier proprietary coding systems while reducing costs by nearly 85%.
Devstral 2 debuts alongside Mistral Vibe, a new command-line interface that enables developers to interact with the model using natural language prompts to execute complex coding operations. A smaller variant, Devstral Small 2—featuring 24 billion parameters—is also available and optimized for local deployment.
Mistral AI aims to offer an open alternative to proprietary coding agents from companies like Google LLC and OpenAI, which often lock users into closed ecosystems. The company positions itself as a genuinely open-weight agentic coding platform.
The Mistral Vibe CLI serves as the central interface, leveraging the underlying Devstral models to translate natural language instructions into real-world code modifications. This system goes beyond generating isolated code snippets—it actively explores, modifies, and implements changes across entire codebases. It operates within any third-party integrated development environment (IDE) or as a standalone terminal tool, equipped with capabilities that allow the model to autonomously edit files, search repositories, manage version control, and execute shell commands.
Mistral AI notes that Devstral 2 can analyze file structures and Git states, granting it “project-aware context” to update dependencies or refactor code across a repository without losing track of ongoing work.
Efficiency is a core focus for Devstral 2. Built on a dense Transformer architecture with 123 billion parameters and a 256,000-token context window, it achieved an impressive 72.2% pass rate on the SWE-bench benchmark, placing it among the industry’s top-performing coding models. Only DeepSeek V3.2 scored higher among openly available models, while frontier systems from Google, OpenAI, and Anthropic PBC also rank above it.
Despite its smaller size, Devstral Small 2 delivers strong performance—scoring 68% on the same benchmark—making it competitive with models five times its scale. Its lightweight design enables efficient execution on consumer-grade hardware like standard laptops, eliminating the latency typically associated with cloud-based AI inference.
Weight, Licensing, and Cost Advantages
According to Mistral AI, Devstral 2’s most compelling advantage lies in its cost efficiency. Priced at $0.40 per million input tokens and $2.00 per million output tokens via its API, it is approximately seven times cheaper than alternatives like Anthropic’s Claude Sonnet 3.5.
Beyond size, Devstral 2 and Devstral Small 2 differ significantly in licensing. Devstral 2 is released under a modified MIT license that imposes revenue-based usage restrictions, while Devstral Small 2 uses the more permissive Apache 2.0 license, allowing unrestricted modification and integration without legal constraints typical of proprietary systems.
Enterprises could easily adopt a hybrid workflow: using the larger 123-billion-parameter Devstral 2 for complex architectural planning, while deploying the leaner 24-billion-parameter Devstral Small 2 for rapid, private code edits that remain within their internal network boundaries.
Mistral AI’s release of Devstral 2 arrives amid intensifying competition in the agentic coding space. Its open-weight approach may appeal to developers wary of vendor lock-in—a trend exemplified by Google’s recent partnership with Replit Inc., which tightly bundles powerful models, IDEs, and cloud infrastructure into a single ecosystem.
Google’s Gemini 3 Pro is deeply integrated with its newly launched AntiGravity IDE, clearly aiming to embed developers firmly within its ecosystem. Meanwhile, Anthropic is reportedly developing a dedicated runtime for its coding agent following its acquisition of the JavaScript toolkit Bun.
