Hcompany has released Holo4, a series of agentic models designed to work across graphical interfaces, code, MCP and APIs. The series includes Holo4-27B, Holo4-35B-A3B and the updated Holotron4 Nano; Holo4-27B and Holo4-35B-A3B are available through the H Models API, while model weights are posted on Hugging Face.
One model, several interfaces
The release targets workflows that move between different software surfaces rather than staying inside a single interface. Hcompany says Holo4 can click and type on a screen, write and run code, and call MCP or API tools. The company describes the same model being used on desktops, websites, Android, code sandboxes and business APIs.
Reported benchmark results
Hcompany reports that Holo4-27B scored 61.7% on OSWorld 2.0, compared with 81.8% for Opus 5.5, while Holo4-35B-A3B scored 30.9%. The company presents the smaller models as substantially cheaper and smaller than leading closed systems, but these are company comparisons rather than independent tests.
The cost and score comparisons use different sources, harnesses and task subsets across models. Hcompany says its Holo4 and selected Qwen results on AutomationBench v1.0.6 came from an internal harness, while some competing results came from public data. Holo4 has not yet been reported on AutomationBench's private set, according to the release.
Training for long workflows
The models were trained with supervised and reinforcement learning on environments and tasks produced partly by Hcompany's Agentic Task Factory. The company says that system has generated about 10,000 tasks across web apps, MCP servers and desktop environments, including settings where the same state is exposed through both a GUI and MCP.
Hcompany also rebuilt its execution harness around feedback from OSWorld 2.0. Its reported changes include longer-lived agent memory for tasks spanning hundreds of steps and access to a shell on the desktop machine. Demonstrations include generating a FreeCAD model, creating a Godot game and completing other professional software tasks, but the source does not provide independent evaluations of those examples.
Weights and follow-up model
Weights are available in BF16, FP8, NVFP4 and 4-bit GGUF formats alongside Holotron4 Nano. Hcompany has also released the trajectories behind its public benchmark scores for replay or download. It says optimized DSpark drafter checkpoints are planned for release in the coming days, so those checkpoints were not yet available in the supplied announcement.
AI-generated editorial illustration; not a documentary photograph.
Source: Holo4: powering generalist computer-use agents
SIGNAL TITANS
Leave a Reply