mini-coder-4b
mini-coder-4b is a 4B parameter model distilled from Qwen 3 Coder 30B A3B. It punches well above its weight, outperforming gpt-oss-120b on SWE-bench Verified Bash only:
| Model | pass@1 | pass@100 |
|---|---|---|
| Qwen 3 Coder 30B-A3B | 33.2 | 67.4 |
| mini-swe-4b | 26.8 | 60.2 |
| gpt-oss-120b | 26.0 | – |
| mini-swe-1.7b | 18.6 | 50.4 |
| SWE-agent-LM 7B | 15.2 | – |
| Qwen 3 4B Instruct 2507 | 4.0 | 25.1 |
It is trained on 400k training trajectories using the lightweight mini-swe-agent scaffolding and the SWE-smith dataset of GitHub issues.
Unlike existing agentic SWE models, the mini-coder models can be post-trained on a single 80GB GPU—or smaller. They work seamlessly with mini-swe-agent, a lightweight, scalable, and developer-friendly agentic framework well-suited for RL fine-tuning. And because they are dense rather than MoE models, they benefit from a more mature fine-tuning ecosystem.
Example usage: Generating SWE-bench trajectories with mini-swe-agent and vLLM
This example shows how to generate SWE-bench trajectories using mini-swe-agent as the agentic scaffolding (recommended) and vLLM as the local inference engine.
First, launch a vLLM server with your chosen model. For example:
vllm serve ricdomolm/mini-coder-4b &
By default, the server will be available at http://localhost:8000.
Second, edit the mini-swe-agent SWE-bench config file located in src/minisweagent/config/extra/swebench.yaml to include your local vLLM model:
model:
model_name: "hosted_vllm/ricdomolm/mini-coder-4b" # or hosted_vllm/path/to/local/model
model_kwargs:
api_base: "http://localhost:8000/v1" # adjust if using a non-default port/address
Create a litellm registry.json file:
cat > registry.json <<'EOF'
{
"ricdomolm/mini-coder-4b": {
"max_tokens": 64000,
"input_cost_per_token": 0.0,
"output_cost_per_token": 0.0,
"litellm_provider": "hosted_vllm",
"mode": "chat"
}
}
EOF
Now you’re ready to generate trajectories! Let's solve the django__django-11099 instance of SWE-bench Verified:
LITELLM_MODEL_REGISTRY_PATH=registry.json mini-extra swebench --output test/ --subset verified --split test --filter '^(django__django-11099)$'
You should now see the generated trajectory in the test/ directory.
Citation
@article{olmedo2026computational,
title={Computational Arbitrage in AI Model Markets},
author={Olmedo, Ricardo and Sch{\"o}lkopf, Bernhard and Hardt, Moritz},
journal={arXiv preprint arXiv:2603.22404},
year={2026}
}