ricdomolm/mini-coder-1.7b: A Compact Code Generation Model
mini-coder-1.7b is a 1.7 billion parameter model, distilled from the larger Qwen 3 Coder 30B A3B. It is specifically engineered for software engineering (SWE) tasks, demonstrating strong performance despite its smaller size.
Key Capabilities & Performance
- Code Generation & Problem Solving: Excels in solving software engineering problems, as evidenced by its performance on the SWE-bench Verified Bash only benchmark.
- Outperforms Larger Models: Achieves a pass@1 score of 18.6 and pass@100 of 50.4 on SWE-bench Verified Bash only, surpassing models like SWE-agent-LM 7B (15.2 pass@1).
- Efficient Training: Trained on 400,000 trajectories using the lightweight mini-swe-agent scaffolding and the SWE-smith dataset of GitHub issues.
- Resource-Friendly Fine-tuning: Unlike many agentic SWE models,
mini-coder-1.7b can be post-trained on a single 80GB GPU or smaller, making it accessible for developers with limited resources. - Seamless Agentic Integration: Designed to work effectively with the mini-swe-agent framework, a scalable and developer-friendly agentic system well-suited for RL fine-tuning.
Ideal Use Cases
- Automated Software Engineering: Best suited for tasks requiring automated code generation, bug fixing, and problem-solving within a software development context.
- Agentic Workflows: Optimized for integration into agentic systems, particularly with the mini-swe-agent, for autonomous code development.
- Resource-Constrained Environments: Its compact size and efficient fine-tuning capabilities make it suitable for deployment and customization on less powerful hardware.