H company published details of Holo4, a new series of agentic models built to operate software through whichever interface is available: graphical interfaces, written code, MCP, or direct API calls. The release includes two sizes, a 27B dense model and a 35B-A3B mixture-of-experts model, both available now on the H Models API and as open weights on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF formats. The company also released Holotron4 Nano, an updated version of Holotron 3 built by applying the same training recipe to NVIDIA's Nemotron 3 Nano Omni model, as part of the NVIDIA Nemotron Coalition.
According to the post, Holo4 is trained to use one model across desktops, the web, Android, code sandboxes, and business APIs, rather than requiring a different model per platform. The company frames this against a common limitation: agentic models optimised for GUI clicking are unusable without a screen, while models built for tool-calling fail against applications that expose no API, and real business tasks often need both.
On OSWorld 2.0, a benchmark for long desktop workflows, Holo4 27B scored 61.7% against 81.8% for Opus 5.5, while Holo4 35B-A3B reached 30.9%. H company says this trails the strongest closed models but at a much lower cost per task, and it has published every benchmark trajectory publicly for replay. The models were trained using supervised and reinforcement learning on environments generated by the company's internal Agentic Task Factory, which has so far produced roughly 10,000 tasks spanning web apps, MCP servers and desktop environments, including hybrid environments exposing the same state through both a GUI and MCP.
H company also rebuilt its training harness, the loop that executes an agent's actions and manages its context over long task sequences, adding a persistent memory across hundreds of steps and direct shell access on the desktop machine. Optimized DSpark drafter checkpoints for faster inference are planned for release shortly.
Top comments (0)