Kronumos/Kronumos-14B-Kairos

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 29, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Kronumos/Kronumos-14B-Kairos is a 14.8 billion parameter language model developed by Tokenectomy Labs, based on the Qwen2.5-Coder-14B-Instruct architecture, with a 32768 token context length. It is specifically optimized for Automated Program Repair (APR) and repository-level patch synthesis, excelling at complex multi-file Abstract Syntax Tree (AST) localization and mathematical invariants. This model is designed for autonomous software engineering, providing fast 1-turn self-healing for algebraic and cross-module AST repairs.

Loading preview...

Kronumos 14B Kairos: High-Capacity Dense Reasoning for Program Repair

Kronumos 14B Kairos is an experimental checkpoint developed by Tokenectomy Labs, specializing in autonomous software engineering for closed-loop program repair and automated bug remediation in production repositories. Built upon the Qwen2.5-Coder-14B-Instruct architecture, this 14.8 billion parameter model is part of the Kronumos Dual-Brain ecosystem, designed for complex algebraic and cross-module Abstract Syntax Tree (AST) repairs.

Key Capabilities

  • Automated Program Repair (APR): Optimized for repository-level patch synthesis and bug remediation.
  • Dense Reasoning: Offers 2x capacity for abstract algorithmic and multi-hop reasoning compared to its 7B predecessor.
  • Native Symbolic Mastery: Excels with deep algebraic trees, yielding high candidate patch rates on challenging repositories like sympy and sphinx.
  • Fast Self-Healing: Achieves 1-turn self-healing for clean patches following feedback from its Rust-based Tokenectomy Sub-Cortex.
  • High Patch Yield: Demonstrates an 85.8%+ yield of syntactically valid, testable git diffs.
  • Extended Context Coherence: Maintains coherence across a 16,384 – 32,768 token context window, handling whole-class inheritance hierarchies without attention degradation.
  • Dual-Brain Architecture: Integrates a 14B Neural Core for hypothesis generation with a Tokenectomy Sub-Cortex (Rust C-ABI) for AST validation and patch anchoring.

Good for

  • Automated Bug Fixing: Ideal for autonomous program repair in production environments.
  • Complex Codebase Maintenance: Suited for tasks involving multi-file AST localization and mathematical invariants.
  • Software Engineering Automation: Useful for developers seeking to automate aspects of code repair and maintenance workflows.
  • Research in APR: Provides a robust platform for exploring advanced automated program repair techniques, as detailed in its Springer Nature preprint.

Finetunes
1 models