Jackrong/Qwopus3.5-4B-Coder
Jackrong/Qwopus3.5-4B-Coder is a 4.5 billion parameter dense transformer model built on the Qwen3.5 architecture, fine-tuned for agentic coding tasks. It excels at code debugging, structured tool use, and reasoning-heavy developer workflows, optimized for local execution with a 32K token context length. The model utilizes Trace Inversion and curriculum SFT to enhance its stability and performance in interactive coding environments.
Loading preview...
Qwopus3.5-4B-Coder: A Compact Agentic Coding Model
Qwopus3.5-4B-Coder is a 4.5 billion parameter model based on the Qwen3.5 architecture, specifically fine-tuned for agentic coding, debugging, and structured tool use. Developed by Jackrong in collaboration with Kyle Hessling, this model is designed for efficient local execution, making it suitable for resource-constrained environments like 16GB laptops.
Key Capabilities
- Structured Debugging: Targets bug localization, minimal patch reasoning, and environment-verified code repair.
- Agent Trace Alignment: Learns from tool-call trajectories with real feedback loops, enhancing interactive agent behavior.
- Local-First Design: Optimized for local deployment, balancing size and reasoning capacity for developer workflows.
- Enhanced Reasoning: Utilizes "Trace Inversion" to reconstruct detailed reasoning traces from compressed summaries, improving learnable Chain-of-Thought (CoT).
- Robust Tool Routing: Achieves 100% on the ToolCall-15 benchmark, demonstrating perfect tool routing stability.
Performance Highlights
Evaluated using benchlocal with Multi-Token Prediction (MTP) n=2:
- Overall Suite Average: 82.0% (vs. 74.0% baseline, +8.0 pp).
- BugFind-15 (Debugging): 71/100 (+19 delta over baseline).
- HermesAgent-20 (Agent Workflow): 64/100 (+3 delta over baseline).
- ToolCall-15 (Tool Routing): 100/100 (perfect score).
Training Methodology
The model employs a three-stage curriculum learning approach, combining Trace Inversion data augmentation with curriculum SFT to gradually expand context length and stabilize formatting. This includes initial stabilization on short reasoning samples, followed by exposure to complex coding and agent traces, and finally reinforcement with longer contexts up to 32K tokens.
Should I use this for my use case?
This model is ideal for local code debugging, small repository tasks, tool-call routing, and structured instruction following where low latency and local execution are critical. It's particularly well-suited for developer workflow assistants. However, as a compact model, it may have limitations with broad world knowledge or highly specialized domain requirements. For contexts beyond 32K tokens, RoPE/YaRN scaling is recommended for stability.