H Company launches Holo4 for GUI, code and tool use

H Company released Holo4, a family of agentic models designed to operate across graphical interfaces, code and external tools. The release includes a 27-billion-parameter dense model and a 35B-A3B mixture-of-experts model. For broader company context, H Company also appeared in a 2026 Startup Battlefield 200 roundup published before the Holo4 release.

H Company says both models can click and type in graphical interfaces, write and run code, and call MCP or API tools. The shared design allows Holo4 to choose among interfaces while working across desktops, the web, Android devices, code sandboxes and business APIs, the company says, rather than requiring a separate model for each environment.

Both models were available through the H Models API at announcement, according to H Company. The company also published weights on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF formats. H Company describes an execution harness that manages actions and context over hundreds of steps, using persistent memory and a shell on the desktop machine.

The company says Holo4 was trained with supervised and reinforcement learning, including tasks from an internal Agentic Task Factory that it claims has produced about 10,000 tasks. That training-scale figure has not been independently corroborated.

H Company reports that Holo4 27B scored 61.7% on OSWorld 2.0, compared with 30.9% for Holo4 35B-A3B. The company notes that task releases, subsets and evaluation harnesses differ, and the scores were not independently reproduced. Its claims that Holo4 lowers cost per task and performs strongly in professional software workflows also rely on company tests and comparisons and were not independently verified.

The announcement also released Holotron4 Nano, which H Company describes as an update to Holotron 3 made by applying its post-training stack to Nemotron 3 Nano Omni.