# H Company releases Holo4 models for software tasks

> Holo4 can work through screens, code and software tools, rather than relying on just one way to interact with a computer. The two models are available through H Company’s API, with downloadable weights as well.

Oossa · 2026-09-28 · https://oossa.com/en/h-company-releases-holo4-models-for-software-tasks

H Company has released Holo4, a series of AI models designed to carry out tasks in software. The models can click and type on a screen, write and run code, or use software tools and APIs. That could let one agent handle several parts of a task without switching to a different model for each type of software.

Holo4 comes in two versions: a 27-billion-parameter model and a 35B-A3B version. Both are available through H Models API, and H Company has also posted model weights for download. The announcement does not list a price for API use.

## One model, several ways to work

H Company says it trained Holo4 with supervised and reinforcement learning across software environments and tasks, including about 10,000 tasks made by its internal Agentic Task Factory. The company says its examples include building 3D models in FreeCAD and making a Pac-Man-style game in Godot. These are demonstrations from the company, not independent assessments of how reliably the models perform on everyday work.

The release also includes Holotron4 Nano, a smaller agent model based on NVIDIA’s Nemotron 3 Nano Omni. H Company says its training process improved the base model on tasks involving graphical interfaces, APIs and coding environments.

## How it scored

On OSWorld 2.0, a test of computer-use tasks, Holo4 27B scored 61.7%. H Company compares that with 81.8% for Opus 5.5. The company says Holo4 costs much less per task, but it does not give a single price in the announcement, and notes that models were tested with different releases and setups.

H Company has published the action-by-action records behind its public benchmark results so people can replay them or download the data. Those records offer a way to inspect what the model did, rather than relying only on a final score.

## The facts

- Holo4 has 27B and 35B-A3B versions.
- Holo4 27B scored 61.7% on OSWorld 2.0; H Company reports 81.8% for Opus 5.5.
- H Company says its Agentic Task Factory has produced about 10,000 tasks.
- Holo4 is available through H Models API, and downloadable weights are offered in several formats.

## Why it matters

For people who use AI to handle office or computer tasks, Holo4’s approach could mean fewer handoffs between tools that work with screens, code and APIs. But the announcement gives no API price, and company demonstrations and benchmark scores do not show how dependable it will be in a particular workplace.

## Sources & references

1. [Holo4: powering generalist computer-use agents](https://huggingface.co/blog/Hcompany/holo4) – Hugging Face, 2026-09-28

Last updated: 2026-09-28