H Company Releases Holo4 Computer-Use Models With Different Self-Hosting Rights

H Company has released Holo4, a pair of AI models designed to move between clicking on screens, writing code and calling software tools. Both are available through an API, and both have downloadable weights. But the choice is not simply between two model sizes: the variant with the stronger reported desktop score has weights restricted to noncommercial use.

One agent, several ways into an application

Holo4 can interact with desktop, web and Android interfaces; write and run code in a sandbox; or call tools through APIs and the Model Context Protocol, which connects AI systems with outside tools. H Company’s release details describe one model handling these methods rather than a separate model for each interface.

An open agent harness passes screenshots and tool results to the model, then carries out its requested actions. H Company says it trained Holo4 on about 10,000 generated tasks across web applications, desktop software and tool servers. Its task-making system checks whether an agent reaches the intended result through the interface.

The deployment choice cuts across the scorecard

The lineup has a 27-billion-parameter dense model and a 35-billion-parameter mixture-of-experts model that uses about 3 billion parameters at a time. Both have a 256,000-token context window through H Models API. But their downloadable weights come with different commercial rights.

The 27B weights carry a noncommercial license, while the 35B-A3B weights permit commercial self-hosting under Apache 2.0. Commercial users can still access 27B through H’s API. Downloading its weights does not grant permission to use them for commercial work.

A lower task price does not settle the comparison

H puts 27B at $1.22 per OSWorld 2.0 task, against a cited $8.48 for Claude Opus 5.5. But Opus scores 81.8% against 27B’s 61.7% in H’s figures. This is not a controlled head-to-head result: the models ran with different agent harnesses and effort settings. Those task costs are not guaranteed customer bills.

H lists API rates per million tokens of $0.40 for input and $3.00 for output on 27B, versus $0.30 and $2.00 on 35B-A3B. Tokens are units processed and generated during a request. Cheaper tokens do not necessarily mean a cheaper completed task.

What the published tests let developers inspect

H also reports a 45.4% score for 27B on AutomationBench, a test of software-tool tasks. Of its 600 public tasks, 480 belong to a split from which H collected training data; that does not mean every task was used in training. H reports 49.3% on the other 120, which it held out. The distinction matters when judging performance on unseen work.

H publishes evaluation trajectories—records of an agent’s steps—and weights in several formats. Developers can inspect those reported runs rather than relying on scores alone. Choosing between the variants still means weighing desktop results against the terms and costs of deployment.