ricdomolm/mini-coder-4b

Model Overview

ricdomolm/mini-coder-4b is a 4 billion parameter model, distilled from the larger Qwen 3 Coder 30B A3B. It is specifically engineered for software engineering tasks, particularly excelling in code generation and problem-solving within a development context.

Key Capabilities & Performance

• Code Generation: Optimized for generating code and solving software engineering problems.
• SWE-bench Performance: Achieves a pass@1 score of 26.8 and pass@100 of 60.2 on the SWE-bench Verified Bash only benchmark, surpassing models like gpt-oss-120b (26.0 pass@1).
• Efficient Training: Trained on 400k trajectories using the lightweight mini-swe-agent scaffolding and the SWE-smith dataset.
• Resource-Friendly: Unlike many agentic SWE models, mini-coder-4b can be post-trained on a single 80GB GPU or smaller, making it accessible for more developers.
• Fine-tuning: Benefits from a mature fine-tuning ecosystem due to its dense model architecture, as opposed to Mixture-of-Experts (MoE) models.

Use Cases

• Agentic Software Development: Works seamlessly with the mini-swe-agent framework for automated problem-solving and trajectory generation.
• Local Inference: Can be deployed with local inference engines like vLLM for efficient, self-hosted code generation.
• Educational & Research: Suitable for experimenting with smaller, yet capable, code generation models on limited hardware.