H Company Releases Holo4 AI Agent Models

H Company has released Holo4, a new family of generalist computer-use models designed for AI agents. These models are capable of interacting with graphical user interfaces (GUIs) by performing clicks and typing on screens, writing code, and calling various tools, including those accessible via MCP or API interfaces. Holo4 aims to address limitations in existing AI agents, which often struggle when graphical interfaces lack APIs or when an agent is solely dependent on screen interaction. The Holo4 family includes two primary sizes: Holo4 27B, a dense model, and Holo4 35B-A3B, a Mixture of Experts (MoE) model with 3 billion active parameters. Both models support a substantial 256,000-token context window and are accessible through the H Models API. Holo4 35B-A3B is released with Apache 2.0 weights, enabling commercial self-hosting, while Holo4 27B weights are under a CC BY-NC 4.0 license, requiring commercial use to be routed through the H Models API. Holo4 is a vision-language model specifically fine-tuned for computer use. The Holo4 27B model is built upon the Qwen1.5-27B architecture, and the Holo4 35B-A3B model is based on Qwen1.5-35B-A3B. Both are integrated with H Company's open-source hai-agents harness, which facilitates the interaction by sending screenshots and tool results to the model and then executing the model's requested actions. H Company positions Holo4 as a cost-effective solution that approaches frontier performance. In benchmarks provided by H Company, Holo4 27B achieved an 85.2% score on OSWorld at a cost of $0.08 per task, significantly lower than its base model Qwen1.5-27B, which scored 84.3% at $0.22 per task. On AndroidWorld, Holo4 27B reached an 85.1% score. For longer workflows, on OSWorld 2.0, Holo4 27B scored 61.7% at $1.22 per task, compared to Claude Opus 5.5's 81.8% score at $8.48 per task, according to H Company's figures. On AutomationBench, Holo4 27B achieved a 45.4% score at $0.05 per task. It is noted that frontier scores are derived from different harnesses and effort levels, and H Company's training data included 480 of the 600 public tasks in AutomationBench. On the remaining 120 held-out tasks, Holo4 27B scored 49.3%. H Company has made all trajectories available on trajectories.hcompany.ai and Hugging Face. The development of Holo4 involved H Company's internal pipelines, referred to as the Agentic Task Factory, which are used to construct environments and verifiable tasks from diverse data sources.