A brand new open-weights AI coding mannequin is closing in on proprietary choices
On Tuesday, French AI startup Mistral AI launched Devstral 2, a 123 billion parameter open-weights coding mannequin designed to work as a part of an autonomous software program engineering agent. The mannequin achieves a 72.2 p.c rating on SWE-bench Verified, a benchmark that makes an attempt to check whether or not AI techniques can remedy actual GitHub points, placing it among the many top-performing open-weights fashions.
Maybe extra notably, Mistral didn’t simply launch an AI mannequin, it launched a brand new improvement app known as Mistral Vibe. It’s a command line interface (CLI) much like Claude Code, OpenAI Codex, and Gemini CLI that lets builders work together with the Devstral fashions immediately of their terminal. The device can scan file buildings and Git standing to take care of context throughout a whole challenge, make modifications throughout a number of information, and execute shell instructions autonomously. Mistral launched the CLI below the Apache 2.0 license.
It’s all the time smart to take AI benchmarks with a big grain of salt, however we’ve heard from workers of the large AI firms that they pay very shut consideration to how properly fashions do on SWE-bench Verified, which presents AI fashions with 500 actual software program engineering issues pulled from GitHub points in in style Python repositories. The AI should learn the problem description, navigate the codebase, and generate a working patch that passes unit assessments. Whereas some AI researchers have famous that round 90 p.c of the duties within the benchmark take a look at comparatively easy bug fixes that skilled engineers might full in below an hour, it’s one of many few standardized methods to match coding fashions.
Concurrently the bigger AI coding mannequin, Mistral additionally launched Devstral Small 2, a 24 billion parameter model that scores 68 p.c on the identical benchmark and may run domestically on shopper {hardware} like a laptop computer with no Web connection required. Each fashions assist a 256,000 token context window, permitting them to course of reasonably massive codebases (though whether or not you think about it massive or small could be very relative relying on general challenge complexity). The corporate launched Devstral 2 below a modified MIT license and Devstral Small 2 below the extra permissive Apache 2.0 license.
