Holo 3.1 puts computer-use agents on your laptop

HCompany released Holo 3.1 this week, a family of computer-use agents built on Qwen that drive web, desktop, and mobile interfaces. There are four sizes: 0.8B, 4B, 9B, and a 35B mixture-of-experts variant with 3B active parameters. The release ships FP8, Q4 GGUF, and NVFP4 quantized builds so the smaller models can actually run on a laptop. NVFP4 inference is reported at 1.74x BF16 throughput while landing within two points of full precision on the listed benchmarks.

The benchmark that jumps out is AndroidWorld. The 35B model goes from 67% to 79.3%, a real step up on tasks like "open this app, type that, tap there" that have stayed stubbornly hard for general assistants. The team frames the release as a push toward agents that work the same way across screen environments and across frameworks, with function-calling protocols added alongside the older JSON-only interface.

What makes Holo 3.1 worth pulling in even if you do not plan to build with it: this is a credible open-weight alternative to closed computer-use APIs, with a small enough footprint to deploy on the device itself. The 9B and below sizes are the ones to look at first. The 35B is interesting, but most production workflows do not need a frontier-class agent to fill in a form.

Why it matters

If you have been waiting for an on-device computer-use model you can actually ship inside an app, this is the most concrete option to test against. Closed APIs still beat it on the hardest tasks, but the gap on routine mobile and web automation is now small enough that the privacy and latency story may matter more.