Details
- H Company launched Holo4 in September 2026, offering 27B dense and 35B-A3B Mixture-of-Experts models through the H Models API and releasing full weights through Hugging Face.
- The models interact with software through graphical interfaces, code, Model Context Protocol (MCP), and APIs, operating across desktops, browsers, Android, code sandboxes, and business systems.
- Holo4 was trained with supervised and reinforcement learning on environments and tasks generated partly by H Company’s Agentic Task Factory; Holotron4 Nano was also released as an updated model.
- On OSWorld 2.0, Holo4 27B scored 61.7% at a reported $1.22 per task, while 35B-A3B scored 30.9% at $0.61; on AutomationBench, they scored 45.4% and 34.5%, respectively.
- H Company reports substantially lower task costs than listed frontier-model comparisons, while the 35B-A3B model is available under Apache 2.0 and the 27B model is designated research-only.
Impact
Holo4 advances a generalist approach to software agents by combining GUI control, coding, and tool calling within one model rather than routing tasks across interface-specific systems. Its lower reported inference costs and open-weight 35B-A3B release could encourage developers to test computer-use agents in enterprise workflows, while putting pressure on vendors to improve long-task reliability and deployment economics over the next 12–24 months.
