Hugging Face has launched Holo4, a new series of agentic models engineered to power generalist computer-use tasks. The series is offered in two sizes, a 27B dense model and a 35B-A3B Mixture of Experts model, both accessible through the H Models API with model weights published on Hugging Face. Additionally, the company released an updated version of Holotron 3 named Holotron4 Nano, which adapts the Nemotron 3 Nano Omni model using the same post-training stack. Holo4 models are designed to interact with software through any available interface, including graphical user interfaces, code, Model Context Protocol, and APIs, avoiding the interface silos common in traditional agentic models. Trained using supervised and reinforcement learning on a vast array of environments and tasks created by the Agentic Task Factory, Holo4 achieves competitive performance on academic benchmarks such as OSWorld 2.0 and AutomationBench while operating at a lower cost than many frontier closed models. Hugging Face also rebuilt its execution harness to improve context management over hundreds of steps, incorporating reliable memory and a desktop shell. The company plans to release optimized DSpark drafter checkpoints to accelerate inference further.
Holo4: powering generalist computer-use agents
Hugging Face has introduced Holo4, a new series of agentic models designed for generalist computer-use tasks across diverse interfaces like GUIs, code, MCP, and APIs. Available in 27B dense and 35B-A3B Mixture of Experts sizes, Holo4 was trained via supervised and reinforcement learning on environments built by the Agentic Task Factory. Alongside Holo4, the company released Holotron4 Nano, an updated model built from Nemotron 3 Nano Omni.
Why it matters
Holo4 demonstrates that training recipes can successfully transform foundation models into cross-interface agentic experts at a lower cost than many closed frontier alternatives. This highlights progress in building flexible agents capable of handling complex real-world software workflows.