AI/ ai agents · open-weight models · coding benchmarks · software repair

Open-Weight Kozuchi Agent Leads Open Models on Bug Repair Tests

A 27B-parameter open-weight coding agent ranks first among open models on two SWE-bench tracks, placing 4th of 42 on Java and 12th of 135 on Python overall.

A research team has released Kozuchi Agent, an open-weight AI system that turns bug reports into code patches without any closed-source model access.

Kozuchi Agent runs on a locally hosted, unmodified Qwen3.5-27B model, with no fine-tuning, using eight-way test-time selection (TTS@8) to pick among candidate patches. On the official SWE-bench Verified evaluator, it resolved 374 of 500 instances. On Multi-SWE-bench, the same agent fixed 41 of 128 Java issues (32.03%), placing fourth of 42 overall and first among strict open-weight entrants; on Python it ranked 12th of 135, again first among open-weight systems. The researchers report per-phase behavior held steady within five percentage points across languages, and that a reusable CI pipeline cut the manual steps needed to run evaluations from five down to one across mixed internal clusters.

The result matters less as a leaderboard placement and more as a demonstration that open, locally run models can compete on structured, tool-heavy engineering tasks, not just chat benchmarks. Teams wary of sending proprietary code to a hosted API get a credible alternative, provided they have the GPU budget for a 27-billion-parameter model plus eight-way sampling per bug. The paper also localizes the remaining gap: mostly patches that pass tests but fail on semantic correctness, plus selection errors, not weaker code editing or model access, which quiets one common excuse for open-weight underperformance.

The paper frames this as favorable among open and local peers by parameter count, not a claim to beat the field outright, a distinction worth keeping since fourth of 42 and 12th of 135 still leave real ground between Kozuchi and the top closed-model entries.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →