Hcompany has released Holo4, a new series of generalist computer-use agents available in two sizes—27B dense and 35B-A3B Mixture of Experts—designed to interact with software through multiple interfaces including graphical user interfaces, code execution, MCP tools, and APIs. The models are trained through supervised and reinforcement learning on professional workflows, enabling them to handle complex tasks that require switching between different software interaction methods. On academic benchmarks, Holo4 demonstrates competitive performance with frontier closed-source models at significantly lower cost. The 27B variant scores 61.7% on OSWorld 2.0 compared to GPT-4 Opus 5.5's 81.8%, while consuming far fewer parameters and requiring substantially less computational resources. Both Holo4 models are available through the H Models API, with full trajectories and datasets released openly on Hugging Face to enable community verification and research. The release addresses a key limitation of existing agentic models, most of which are optimized for single interfaces—GUI-focused models struggle without visual screens while code-centric agents lack API integration. Holo4's unified architecture supports real-world business workflows spanning desktops, web applications, Android devices, and enterprise APIs, eliminating the need to select different models for different platforms.