Holo4 Launches Generalist Agentic Models for Multi-Interface Workflows

H Models has unveiled Holo4, a new series of generalist agentic models designed to interact with software across multiple interface types, including graphical user interfaces, code environments, MCP servers, and traditional APIs. The series launches with two configurations: a 27B parameter dense model and a 35B-A3B Mixture of Experts variant. Both are immediately available via the H Models API, with open-weight releases in BF16, FP8, NVFP4, and 4-bit GGUF formats hosted on Hugging Face. Unlike specialized agentic models trained for single interface types, Holo4 is engineered to adapt dynamically to the most efficient interaction method for a given task. The model operates seamlessly across desktop operating systems, web browsers, Android devices, code sandboxes, and enterprise APIs using a unified architecture. This cross-platform consistency eliminates the need for platform-specific model selection, streamlining deployment for complex business workflows. The training foundation relies on H Models Agentic Task Factory, which generates approximately 10,000 interactive environments and verifiable tasks from software documentation and real-world application screenshots. Holo4 underwent supervised learning and reinforcement learning across these hybrid environments. Engineers also overhauled the model execution harness to support reliable multi-step memory management and integrated a native desktop shell, significantly improving stability during extended agentic sequences. In benchmark evaluations, Holo4 demonstrates strong competitive positioning. On the OSWorld 2.0 desktop control benchmark, the 27B variant achieves a 61.7% success rate, trailing only leading closed-source models, while the 35B-A3B MoE model reaches 30.9%. On AutomationBench for API automation, Holo4 similarly matches frontier performance metrics. Notably, these results are attained with orders of magnitude fewer parameters and substantially lower token costs. H Models has also committed to full transparency by open-sourcing all training trajectories used to generate benchmark scores. Performance demonstrations highlight the model's practical utility. In controlled testing against its Qwen3.8 27B base, Holo4 successfully completed complex 3D modeling tasks in FreeCAD and autonomous game development in Godot with fewer tool calls and significantly reduced token consumption. The model's ability to combine GUI navigation, script execution, and API integration enables it to manage intricate, multi-stage software operations without manual intervention. Alongside the main series, H Models released Holotron4 Nano, an optimized agentic variant built by applying the company's post-training stack to the NVIDIA Nemotron 3 Nano Omni model. This adaptation delivers marked improvements in GUI workflow accuracy and multi-interface reasoning. To further enhance inference speed, the company plans to deploy optimized DSpark drafter checkpoints in the near future. Holo4 represents a strategic shift toward cost-efficient, interface-agnostic AI agents capable of operating across fragmented enterprise software ecosystems.