H Company open-sources Holo4, a computer-use VLM family for GUI agents

H Company published the Holo4 family on 28 September in a Hugging Face blog post, describing it as a state-of-the-art vision-language model line for computer and mobile use, tool calls and code. Rather than a single checkpoint, the release covers three different base architectures and several quantisation formats, with weights published under the Apache 2.0 licence.

The family has three members. Holo4-27B uses a Qwen3.8 dense architecture, Holo4-35B-A3B is built on a Qwen3.5 mixture-of-experts design, and Holotron4-30B-A3B follows the NemotronH Nano Omni architecture. All three ship in BF16, while the two Holo4 sizes also come in FP8, NVFP4 and a Q4 GGUF build for local inference.

What turns the models into agents is the harness rather than the weights alone. Per the model card, hai-agents feeds screenshots and tool results to Holo4, executes the clicks, typing, code and tool calls the model requests, then sends the results back; the documentation also covers function calling, element localisation and document OCR. The configured context length is 262,144 tokens.

On performance, H Company reports 85.2% on OSWorld for Holo4-27B at $0.08 per task, and on OSWorld 2.0 scores of 61.7% for Holo4-27B and 30.9% for Holo4-35B-A3B at $1.22 and $0.61 per task respectively. On AutomationBench the same two models score 45.4% and 34.5%.

The company also evaluated the models on Agentic Task Factory, a set of held-out business workflows spanning web, desktop and MCP tools, and released every evaluation trajectory as an open dataset, which lets outside teams audit exactly where the models succeed and where they loop.

The more consequential claim is the cost curve: a 27B dense model that matches or beats larger systems at a few cents per task would move computer-use agents out of demos and into routine back-office work.

Independent reproduction is the thing to watch. Open trajectories make failure modes auditable, the Apache 2.0 licence removes most legal friction for commercial pilots, and the shipped quantised variants lower the deployment bar — but screen-level control also raises unanswered questions about permissions, auditing and who owns a mistaken click.