A debugging-only model claiming 80% on SWE-bench Lite

What it does

The repository houses research artifacts, benchmark results, and a detailed architecture paper for Chronos, a language model built exclusively for autonomous debugging. Unlike general-purpose models that reportedly collapse from ~70% success on code generation to under 15% on debugging tasks, Chronos is trained on 42.5 million debugging examples and optimized for multi-file patch generation, root-cause analysis, and iterative fix validation. The actual model is proprietary and slated for release through Kodezi OS in early 2026; the code here is strictly documentation and evaluation data.

The interesting bit

The README leans heavily into the “debugging gap”—the idea that writing code and fixing it require fundamentally different skills, and that bigger context windows alone do not help. Chronos attacks this with a seven-layer stack that includes Adaptive Graph-Guided Retrieval for multi-hop code traversal and Persistent Debug Memory for repository-specific learning across sessions. The reported numbers are aggressive: 80.33% on SWE-bench Lite, 89% root-cause accuracy, and a claimed 15.7× improvement over baselines on million-line codebases.

Key highlights

• Claims 80.33% on SWE-bench Lite (241/300 instances), with a reported 20-point absolute lead over the next entry.
• Generalist models like GPT-4.1 and Claude 4.5 Sonnet are listed at roughly 13–14% on the same debugging benchmark, despite scoring far higher on code-generation tasks.
• Architecture centers on a “debug-tuned LLM core” fed by an adaptive retrieval engine and a persistent memory layer that reportedly improves from 35% to 65% success as it learns a repository’s bug patterns.
• Custom MRR benchmark (5,000 scenarios) claims 67.3% autonomous debug success and 92% retrieval precision.
• Performance scales inversely with codebase size in absolute terms (59.7% for >1M LOC), but the relative gap over baselines allegedly widens dramatically as repositories grow.

Caveats

• The model itself is not open source; this repository contains only papers, benchmarks, and marketing materials.
• Several baseline comparisons in the README mix different evaluation setups (e.g., “Bash Only” vs full agentic runs), making direct apples-to-apples comparison harder than the tables suggest.
• The 89% human-preference and 40% time-reduction figures lack detailed methodology in the visible README sections.

Verdict

Worth reading if you are researching automated program repair or evaluating vendor claims in the dev-tools space. Skip it if you are looking for an open weights model or a tool you can run locally today.