Microsoft's 4B hybrid computer-use agent scores 66.0% on OSWorld-Verified
A Microsoft 4B hybrid agent that can both click and write code scores 66.0% on OSWorld-Verified, ahead of 27B and 30B agents, per a post.
Microsoft's 4B hybrid computer-use agent, which can both click through an interface and write code, scores 66.0% on OSWorld-Verified, according to a social media post from @HuggingPapers. The post says that result is ahead of agents with 27B and 30B parameters. The "hybrid" label refers to the agent's ability to click and to write code.
The figure is a claim relayed from a post, and nothing independent confirms it. The post does not say how the model decides when to click and when to write code, which 27B and 30B agents it was compared against, or what settings were used in the benchmark run. Without those details, the 66.0% score cannot be checked against the run itself.
What changed
Type: Research · Applies to: Microsoft's 4B hybrid computer-use agent
Why it matters
A small agent that performs well on computer-use benchmarks could lower the cost of automating desktop workflows.
Original source
We have no link to the original source for this story yet.
Where we saw it
Latent Space
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works