Brief, mostly facts

OS-Helix unifies GUI actions and code execution, hits 66.44% OSWorld

An end-to-end trained desktop agent reaches 66.44% on OSWorld-Verified with its 27B RL model. Models and dataset are published on Hugging Face.

Read the source at huggingface.co

What the research found

OS-Helix is an end-to-end trained desktop agent unifying GUI actions and code execution, reaching 66.44% on OSWorld-Verified with OS-Helix-27B-RL, with 27B, 35B-A3B and RL models and a dataset already published on Hugging Face.

Claims and how well the evidence supports them

OS-Helix-27B-RL reaches 66.44% success on OSWorld-Verified 99%
OS-Helix-27B and 35B-A3B achieve 35.0% and 33.7% relative improvements over Qwen3.5 hybrid baselines 98%
Step-balanced GSPO aggregates importance weight by action step rather than raw tokens 99%
Model weights for OS-Helix-27B, 35B-A3B and 27B-RL are published on Hugging Face 86%

Evidence

huggingface.co https://huggingface.co/papers/2610.22234
huggingface.co https://huggingface.co/YYangzzzz/OS-Helix-27B
huggingface.co https://huggingface.co/YYangzzzz/OS-Helix-35B-A3B
huggingface.co https://huggingface.co/YYangzzzz/OS-Helix-27B-RL
huggingface.co https://huggingface.co/datasets/YYangzzzz/OS-Helix-Dataset

Actionability rated medium.