DJLougen/Qwable-5-27B-Coder

Qwable-5-27B-Coder: An Agentic Coding Model

Qwable-5-27B-Coder is a 27 billion parameter model developed by DJLougen, built upon the Qwen3.6 architecture. It is specifically fine-tuned for agentic coding behavior, focusing on the practical, iterative aspects of software development. The model's training involved Claude Fable 5 traces and Kimi 2.7 Coder traces, emphasizing a workflow of inspection, decision-making, editing, verification, and recovery.

Key Capabilities

• Repository Navigation: Designed to work within real code repositories, not just isolated snippets.
• Error Correction: Translates failing command outputs into actionable patches.
• Constraint Management: Maintains project constraints across multi-step coding tasks.
• Tool-Friendly Output: Produces implementation-oriented answers suitable for tool integration.
• Long Context Handling: Processes extensive engineering prompts, including logs, diffs, and stack traces.
• Action-Oriented: Biases towards concrete edits, commands, and verification over generic advice.

Good For

• Coding Agents: Ideal for automated coding workflows and agentic systems.
• Repository Work: Tasks involving navigating and modifying existing codebases.
• Terminal Workflows: Scenarios requiring interaction with terminal feedback and command-line tools.
• Tool-Use Style Chat: Applications where the model needs to generate tool-friendly responses and function calls.

While the model supports an image-text-to-text pipeline due to its base, its fine-tuning is specifically for coding behavior. Public benchmark scores are not yet available, but early maintainer runs indicate performance gains over the base model on private coder benchmarks. Users should note that this is a large BF16 checkpoint, requiring significant resources or quantization for local deployment.