#

Key takeaways

• H Company announced Holo4 on September 28, 2026.

• The release includes a 27B dense model and a 35B-A3B mixture-of-experts model.

• Holo4 is available through the H Models API, with weights also on Hugging Face.

• The system combines graphical interfaces, code execution, MCP, and API tools.

• The source reports benchmark results, but it also notes clear comparison limits.

H Company introduces Holo4

H Company announced Holo4 through a team article on Hugging Face. The release presents Holo4 as a generalist computer-use model series. It also introduces Holotron4 Nano as a successor in the smaller-model line.

The source says the series includes two main models. One is a dense model with twenty-seven billion parameters. The other is a mixture-of-experts model described as 35B-A3B.

What Holo4 is designed to do

Holo4 is designed to combine several tools in one workflow. The source lists graphical interfaces, code execution, Model Context Protocol, and API tools. It also says the supported environments include desktops, web applications, Android, and coding sandboxes.

This makes the release broader than a single-task model. It is framed as a system for computer use across different environments. The source does not claim reliable unattended work, so that limitation remains important.

Training and task sources

H Company says supervised learning and reinforcement learning used environments and tasks from its Agentic Task Factory. The blog reports approximately ten thousand tasks. These tasks span web, desktop, and MCP environments.

The source also says published benchmark trajectories can be inspected or downloaded. That gives readers a way to review the reported evaluation material. It also suggests the release is meant to be examined, not only consumed as a product claim.

Reported benchmark results

The company reports OSWorld 2.0 results for both main models. Holo4 27B is given a score of 61.7 percent. The mixture-of-experts variant is reported at 30.9 percent.

The source adds an important caution. Benchmark versions, harnesses, and task subsets differ. Because of that, comparisons must keep their limitations in view. The company also says improved memory and a shell on the desktop machine were important harness changes.

Access and packaging

The release says both models are available through the H Models API. It also says model weights are published on Hugging Face in multiple formats. That means the release is not limited to one access path.

The source also mentions additional inference-acceleration checkpoints planned for later. No further details are provided. So the release should be read as an active model rollout, not a final endpoint.

Operational considerations

The source does not provide a guarantee of reliable unattended work. That is a direct limitation of the release. It also does not verify any Moroccan commercial deployment or country-specific service access.

For readers, the main operational point is simple. Reported benchmark scores help with comparison, but they do not remove evaluation needs. The source itself stresses that harness changes and task subsets affect interpretation.

Morocco relevance

The source reports no verified Morocco-specific deployment or service access. The only safe takeaway is conditional and global: if a team evaluates similar computer-use models, it should test them in its own workflow before adoption.

Bottom line

Holo4 is presented as a multi-environment computer-use model series with API access and published weights. The release combines model packaging, task-based training, and benchmark reporting in one announcement. At the same time, the source keeps the limits clear, especially around benchmark comparison and unattended reliability.