Holo4-27B is a vision-language model for computer use from H Company, built on Alibaba's Qwen3.8-27B. It is a 27B dense model with a configured maximum context of 262,144 tokens, and it targets web, desktop and mobile environments through graphical interfaces, code and tool calls. It is meant to be used with H Company's hai-agents harness, which sends screenshots and tool results to the model, then executes the clicks, typing, code and tool calls it requests and returns the results; the harness can also give the model access to application tools and code execution. The card also points to H Models API documentation for function calling, element localization and document OCR. On the card's own evaluations, Holo4-27B scores 85.2% on OSWorld at $0.08 per task, 61.7% on OSWorld 2.0 at $1.22 per task, and 45.4% on AutomationBench at $0.05 per task. The card says Holo4 models improve significantly over their Qwen base models, and also reports results on Agentic Task Factory, a set of held-out business workflows across web, desktop and MCP tools; the full benchmark comparison is an image and a blog post, not text. The Holo4 family also includes Holo4-35B-A3B (Qwen3.6 MoE) and Holotron4-30B-A3B (NemotronH Nano Omni), and this BF16 checkpoint has FP8, NVFP4 and Q4 GGUF companion releases. H Company shares all agent evaluation trajectories in an open dataset (Hcompany/trajectories). The weights are released under CC BY-NC 4.0, which allows non-commercial use only, while the Qwen3.8 base is Apache-2.0 and the repository includes both licenses.