H Company · 2026-06-02 · major
H Company Open-Sources Holo3.1 — Computer-Use VLM Family at 0.8B, 4B, 9B, and 35B-A3B Hits 74.2% OSWorld, Jumps AndroidWorld From 67% to 79.3%, and Ships FP8, NVFP4, and Q4 GGUF Checkpoints Under Apache 2.0
H Company ships Holo3.1, an Apache 2.0 computer-use VLM family at 0.8B/4B/9B/35B-A3B. The 35B model hits 74.2% OSWorld and 79.3% AndroidWorld; FP8/NVFP4/Q4 GGUF quantizations drop step time from 6.8s to 3.3s on DGX Spark.
An Apache 2.0 computer-use VLM family that runs locally on consumer GPUs and now drives mobile as well as desktop.
Key specs
| License | Apache 2.0 |
|---|---|
| Smallest model | 0.8B |
| Largest model | 35B-A3B |
| Osworld | 74.2% |
| Android world (35 b) | 79.3% |
| Android world prior (35 b) | 67% |
| Dgx spark step time | 6.8s → 3.3s |
What is it?
Holo3.1 is the second iteration of H Company's open-weights vision-language model family for computer-use agents — models that look at a screen and decide where to click or type. The release ships four sizes (0.8B, 4B, 9B, and a 35B Mixture-of-Experts with 3B active parameters) finetuned from the Qwen 3.5 base, all under Apache 2.0.
How does it work?
Each model takes a screenshot plus an instruction and emits structured actions (click, type, scroll, swipe) with native function-calling so the same checkpoint plugs into different agent frameworks without scaffolding. Training extended coverage from web/desktop to Android, lifting the 35B's AndroidWorld score from 67% to 79.3% and the 4B variant from 58% to 72%. FP8, NVFP4, and Q4 GGUF checkpoints lose ~2 OSWorld points vs. the BF16 master while NVFP4 delivers 1.74× the throughput of BF16, dropping average step time on a DGX Spark from 6.8s to 3.3s.
Why does it matter?
Open computer-use models have lagged behind closed assistants because of compute and latency. A 4B variant that runs on a 12GB consumer GPU and a 35B-A3B that comfortably fits on a single workstation move local desktop and phone automation out of demo territory. Enterprises that cannot ship continuous screenshots to a cloud API now have a credible offline path.
Who is it for?
Agent builders, RPA teams, and on-device automation researchers
Try it
pip install holo-sdk && huggingface-cli download Hcompany/Holo-3.1-4B