CUGA harness reaches 81% AppWorld task completion with GLM-5.3-flash
Overview
Hugging Face reports that its CUGA agent harness completed 81.1% of 90 AppWorld tasks using the open-weight GLM-5.3-flash model, compared with 46.7% for Hermes and 30.0% for Deep Agents.
The company says all three setups ran the same model and task set. Tasks passed by all three variants were 66.7% for CUGA, 26.7% for Hermes and 6.7% for Deep Agents.
The authors state that the setups were not identical and that the sample covers only one benchmark. The results come from Hugging Face's own blog post.
Written by AI from the articles below · updated Nov 3, 11:25 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- Hugging Face BlogOfficialCUGA harness lifts open-weight model's AppWorld task completion to 81% versus 30–47% for rivals
AIHugging Face compared CUGA, Hermes and Deep Agents on AppWorld using the same GLM-5.3-flash model and 90 tasks, with task completion at 81.1% for CUGA, 46.7% for Hermes and 30.0% for Deep Agents. Scenarios where all three variants passed were 66.7% for CUGA, 26.7% for Hermes and 6.7% for Deep Agents. The authors note the setups were not identical and the sample covers one benchmark.
Heat Trend
Not enough continuous observations to show a trend yet.