Mistral’s Devstral: The Open-Source AI That’s Outcoding Google and DeepSeek

A 24B-Parameter Powerhouse for Developers—and It Fits on Your Desktop

Paris-based Mistral AI just dropped a bombshell in the coding AI space. Their new model, Devstral, developed with All Hands AI, isn’t just another incremental upgrade—it’s beating Google’s Gemma 3 27B and DeepSeek’s V3 on the rigorous SWE-Bench Verified benchmark. Even better? It’s fully open-source (Apache 2.0 license) and designed to run locally on a single Nvidia RTX 4090 or a 32GB RAM Mac. For developers tired of cloud latency and vendor lock-in, this changes everything.

“Devstral isn’t just about raw performance—it’s built for real-world workflows: navigating codebases, editing files, and automating software agents,” says a Mistral engineer.

Priced aggressively at $0.1 per million input tokens and $0.3 per million output tokens, Devstral undercuts many closed-source rivals while offering superior speed. Available on Hugging Face and Mistral’s API, it’s already attracting attention from enterprises like BNP Paribas and AXA—clients who’ve bet big on Mistral since its 2023 founding. With €1.1 billion (~$1.24 billion) in funding, the startup is moving faster than expected: Devstral is their third product launch in a month, following Mistral Medium 3 (a general-purpose model) and Le Chat Enterprise (a corporate chatbot).

At 24 billion parameters, Devstral is technically a “research preview,” hinting at Mistral’s plans for an even larger agentic coding model soon. But don’t let the label fool you—its performance on real-world tasks (like patching bugs or generating deployment scripts) suggests it’s ready for prime time. For indie devs and Fortune 500 teams alike, the era of open, local, and affordable coding AI might finally be here.