Skywork-SWE-32B is a code agent model developed by Skywork AI, specifically designed for software engineering (SWE) tasks. It demonstrates strong performance across several key metrics:
When incorporated with test-time scaling techniques, the performance further improves to 47.0% accuracy, surpassing the previous SoTA results for sub-32B parameter models.
We clearly demonstrate the data scaling law phenomenon for software engineering capabilities in LLMs, with no signs of saturation at 8209 collected training trajectories.
We also introduce an efficient and automated pipeline for SWE data collection, culminating in the creation of the Skywork-SWE dataset---a large-scale, high-quality dataset featuring comprehensive executable runtime environments. Detailed descriptions are available on our technical report.
Data Scaling Law for Pass@1 Accuracy on Qwen2.5-Coder-32B-Based LLMs Using the OpenHands v0.32.0 Code Agent Framework. Skywork-SWE-32B significantly outperforms previous Qwen2.5-Coder-32B-based LLMs, achieving the highest pass@1 accuracy without using verifiers or multiple rollouts.
With the incorporation of test-time scaling techniques, Skywork-SWE-32B further improves to 47.0% accuracy, surpassing the previous SoTA results for sub-32B parameter models.
# Install vLLM version 0.9.0.1.
# For example, if your CUDA version is 12.8, use the following command:
pip install vllm==0.9.0.1 --extra-index-url https://download.pytorch.org/whl/cu128
Since our model has 32 billion parameters and supports a 32K context length, we recommend launching the model server with at least 2 GPUs equipped with sufficient VRAM to ensure efficient inference.
If you want to run the OpenHands agent with test-time scaling techniques (a Best-of-N method based on the critic model), please refer to the blog for detailed instructions. You will need to switch to the feature/llm-critic branch and deploy the critic model accordingly. Additionally, you need to add the following parameters into the configuration file:
If you use Skywork-SWE in your research, please consider citing our work using the following BibTeX entry:
@misc{zeng2025skyworksweunveilingdatascaling,
title={Skywork-SWE: Unveiling Data Scaling Laws for Software Engineering in LLMs},
author={Liang Zeng and Yongcong Li and Yuzhen Xiao and Changshi Li and Chris Yuhao Liu and Rui Yan and Tianwen Wei and Jujie He and Xuchen Song and Yang Liu and Yahui Zhou},
year={2025},
eprint={2506.19290},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2506.19290},
}
@misc{skywork-swe,
title={Skywork-SWE: Unveiling Data Scaling Laws for Software Engineering in LLMs},
author={Liang Zeng, Yongcong Li, Yuzhen Xiao, Changshi Li, Chris Yuhao Liu, Rui Yan, Tianwen Wei, Jujie He, Xuchen Song, Yang Liu, and Yahui Zhou},
howpublished={\url{https://quixotic-sting-239.notion.site/eb17f379610040ceb54da5d5d24065bd}},
note={Notion Blog},
year={2025},
}
Files
29 files · 65.5 GBWeights 65.5 GB (safetensors). Sizes and sha256 from the signed manifest.
import os
os.environ["HF_ENDPOINT"] = "https://freeweights.allternit.com"
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Skywork/Skywork-SWE-32B", revision="f39ff4903266ec098e7336a82caac134f7a67355")
model = AutoModelForCausalLM.from_pretrained("Skywork/Skywork-SWE-32B", revision="f39ff4903266ec098e7336a82caac134f7a67355", device_map="auto")
