ProductFR

Devstral 2

Mistral's 123B-parameter agentic coding model, built for multi-file software engineering tasks.

Last refreshed: 17 May 2026 · Appears in 1 active topic

Key Question

Is Devstral 2 the first open-weight coding agent to match closed US frontier models?

Timeline for Devstral 2

#529 Apr

Bundled in Le Chat Enterprise enterprise stack

European Tech Sovereignty: Mistral ships Le Chat Enterprise and Medium 3.5

View full timeline →

Follow European Tech Sovereignty →

Common Questions

What is Mistral Devstral 2 and how does it differ from the original Devstral?

Devstral 2 is Mistral's second-generation agentic coding model at 123B parameters, up from the original. It achieves 72.2% on SWE-bench Verified, supports 256K-token context, and is paired with the new Vibe CLI for terminal-based agentic workflows.Source: Mistral AI

How does Devstral 2 perform on SWE-bench compared to Claude and GPT-4?

Devstral 2 scores 72.2% on SWE-bench Verified, placing it among the top coding agents globally. Mistral claims it outperforms several larger proprietary models on real-world repository tasks at 7x lower cost than Claude Sonnet.Source: Mistral AI / VentureBeat

Can Devstral 2 be run locally or does it require Mistral's cloud?

Devstral Small 2 (24B) is designed for local and laptop deployment under Apache 2.0. The full 123B model runs best on server infrastructure but both are available on Hugging Face for self-hosting.Source: Mistral AI

Background

Devstral 2 is Mistral AI's second-generation developer agent model, launched in December 2025 alongside the Mistral Vibe CLI. With 123 billion parameters and a 256K-token context window, it is purpose-built for agentic software engineering: autonomously navigating codebases, writing tests, resolving bugs across multiple files, and submitting pull requests. A smaller companion, Devstral Small 2 (24B parameters), targets local and laptop deployment. Mistral announced both models as part of its broader Le Chat Enterprise push .

On the SWE-bench Verified benchmark, a standard for evaluating real-world software engineering in existing repositories, Devstral 2 achieves 72.2% and Devstral Small 2 achieves 68.0% — the highest score of any open-weight model at its parameter count, outperforming several 70B-class competitors. Devstral 2 is released under a modified MIT licence; Devstral Small 2 ships under Apache 2.0. Both are available via Mistral's API and Hugging Face. Mistral claims Devstral 2 is up to 7x more cost-efficient than Claude Sonnet on real-world agentic coding tasks.

Devstral 2 is strategically significant for the European AI stack: it is the only open-weight agent-class coding model with a European company behind it. For regulated industries requiring on-premises agentic AI — banking, defence, healthcare — an EU-origin open-weight model addresses both data-residency and AI-Act compliance concerns that US-origin closed models cannot satisfy.

Source Material

Mistral launches powerful Devstral 2 coding model | VentureBeat Introducing: Devstral 2 and Mistral Vibe CLI | Mistral AI

How the World Sees Them

France

Flagship demonstration that a Paris-based lab can produce frontier-class agentic AI; fits squarely into the government's AI sovereignty narrative.

Germany

Open-weight release with Apache 2.0 (small model) allows corporate self-hosting; particularly attractive to large industrial firms with strict IP and data-residency policies.

United States

Anthropic's Claude Code and GitHub Copilot Workspace are the direct competitors; Mistral's cost-efficiency claim of 7x versus Sonnet is a direct commercial challenge.

Enterprise buyers (pan-EU)

The open licence and self-hostable design make Devstral 2 the only EU-origin option for regulated environments requiring agentic AI without US cloud dependency.