Holo4: open-weight computer-use agents for screen, code, and API
H Company's 27B dense and 35B MoE drive desktop, web and Android — ~79% fewer tokens than their base, with open weights and GGUF quants out today.
Copy markdown
One model, any interface
Holo4 clicks and types on screens, writes and runs its own code, and calls MCP or APIs — using whichever fits the task across desktop, web, Android, and code sandboxes. Two open sizes ship today: a 27B dense model and a 35B-A3B MoE. What you can build with it: self-hosted back-office automation, cross-platform QA bots, or internal tools that need auditable action logs.
79% fewer tokens than its base
On a Godot "build Pac-Man" task, Holo4-27B finished in 2.4M tokens over 68 tool calls versus 11.4M tokens and 197 calls for its base model (Qwen 3.8 27B) — about 79% fewer tokens and 65% fewer tool calls. For long agent loops, that gap is the difference between a run you can afford and one you can't.
The OSWorld gap, stated plainly
It isn't frontier: Holo4-27B scores 61.7% on OSWorld 2.0 against Opus 5.5's 81.8%. The trade is openness — weights you can self-host, cheaply, with full control of the action log, for teams that can't send their screens to a closed API.
Weights are up, in every quant
Grab it now on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF — the 4-bit 27B is small enough to run locally. Commercial use is allowed through the managed H Models API or self-hosting; check the model card for exact license terms before you ship.