Kronumos 14B Kairos (15B)
High-Capacity Dense Reasoning for Complex Multi-File Repository Repair
Official Documentation & Interactive Benchmarks • Research Website • Springer Nature DOI • GitHub Core
Dense reasoning checkpoint optimized for multi-file AST localization and mathematical invariants.
⚡ Kronumos 14B Kairos (Experimental Checkpoint)
[!NOTE]
🧪 SOVEREIGN EDGE APR ARCHITECTURE
Kronumos 14B Kairos is the sovereign mid-scale model of the Kronumos Dual-Brain ecosystem, engineered to tackle complex algebraic, cross-module AST repairs (
sympy,sphinx) with fast 1-turn self-healing. For cloud planetary scale, see Kronumos Aion. For high-velocity offline edge, see Kronumos Kairos.
📌 Model Overview
Kronumos 14B Kairos is an autonomous software engineering model specialized for closed-loop program repair and automated bug remediation on production repositories.
- Base Architecture: Qwen2.5-Coder-14B-Instruct (bfloat16)
- Primary Domain: Automated Program Repair (APR) & Repository-Level Patch Synthesis
- Target Benchmark: Princeton SWE-bench Verified (500 instances)
- Sub-Cortex Runtime: Deterministic AST surgical engine written in Rust (Tokenectomy C-ABI 5µs)
- Author & Research Lead: Muhammad Naufal Daffa (ORCID: 0009-0000-7909-4916) | Tokenectomy Labs
- Official Preprint: Springer Nature Research Square — DOI: 10.21203/rs.3.rs-11205335/v1
- GitHub Repository: Tokenectomy-Labs/Kronomus
🚀 Key Architectural Advantages Over Kronumos 7B
| Capability / Metric | Kronumos 7B | Kronumos 14B Kairos | Concrete Impact |
|---|---|---|---|
| Parameter Scale | 7.6 Billion | 14.7 Billion | 2x capacity for abstract algorithmic and multi-hop reasoning |
| Symbolic & Math Repos | Struggled with deep algebraic trees | Native Symbolic Mastery | High candidate patch yield on challenging repos (sympy, sphinx) |
| AST Self-Healing Velocity | Required 3–4 turns to recover syntax | Fast Turn 2 Self-Healing | Recovers clean patches in 1 turn following Sub-Cortex compiler feedback |
| Candidate Patch Yield | ~72.4% yield | 85.8%+ yield | Drastically higher rate of syntactically valid, testable git diffs |
| Context Window Coherence | 8,192 tokens | 16,384 – 32,768 tokens | Holds whole-class inheritance hierarchies without attention degradation |
| Hardware Marginal Cost | ~0.12 Colab units / instance | ~0.19 Colab units / instance | Remains strictly cost-bounded (~$5 total for 500 instances on A100) |
⚠️ Important Usage Notes
Kronumos is designed as a cybernetic dual-brain architecture:
- The 14B Neural Core generates candidate AST mutation hypotheses and targeted SEARCH/REPLACE blocks.
- The Tokenectomy Sub-Cortex (Rust) validates full-file AST syntax, verifies relative indentation, and anchors patches to prevent dirty diffs.
git clone https://github.com/Tokenectomy-Labs/Kronomus.git
cd Kronomus
python3 scripts/kaggle_kronumos_runner.py --model_id NadevA23/Kronumos-14B-Kairos --bfloat16
📖 Citation
@article{daffa2026kronumos2,
author = {Muhammad Naufal Daffa},
title = {Kronumos 2 Kairos: Cost-Bounded Automated Program Repair via Dual-Brain Cybernetic Sub-Cortex on SWE-bench Verified},
journal = {Research Square},
year = {2026},
doi = {10.21203/rs.3.rs-11205335/v1},
url = {https://doi.org/10.21203/rs.3.rs-11205335/v1}
}
Maintained by Tokenectomy Labs.