The news
French AI developer H Company released Holo4, a family of models for computer-use agents, on Monday, September 28, 2026. The Holo4 computer-use model is designed to operate software through its screen, its code or its APIs, so one model can handle mixed business tasks.
Holo4 comes in two sizes, according to the company's announcement on Hugging Face: a 27-billion-parameter dense model and a 35-billion-parameter mixture-of-experts model, labeled 35B-A3B, that uses about 3 billion parameters at a time. H said both are available through its H Models API, and the weights are downloadable from Hugging Face in several formats, including a compressed 4-bit version. The Register described them as open-weight models.
Computer use means an AI agent that works a screen the way a person does: pointing, clicking, scrolling and typing. H says most agent models are trained for only one kind of interface, so a model that relies on screens is stuck without one, and a model that relies on tool calls is stuck when an application has no API. Holo4, the company says, runs on desktops, the web, Android, code sandboxes and business APIs as the same model, called the same way.
The models are built on Alibaba's Qwen models, Qwen3.8 27B and Qwen3.6 35B-A3B, and were trained with supervised and reinforcement learning, H said. The company said its internal Agentic Task Factory, which builds practice environments and tasks from documentation, has produced about 10,000 tasks across web apps, MCP servers and desktop environments. MCP, or Model Context Protocol, is a standard way to connect AI models to outside tools.
On OSWorld 2.0, a benchmark for desktop control, H reported that Holo4 27B scored 61.7% and Holo4 35B-A3B scored 30.9%, compared with 81.8% for Opus 5.5. In the same chart, H listed Opus 5 at 70.2% and GPT-5.6 Sol at 66.2%, noting that task sets and testing harnesses differ between sources. H also released Holotron4 Nano, an updated model built on Nvidia's Nemotron 3 Nano Omni.
The numbers
- Holo4 model sizes
- 27B dense; 35B-A3B mixture-of-experts
- OSWorld 2.0 score, Holo4 27B (company-reported)
- 61.7%
- OSWorld 2.0 score, Holo4 35B-A3B (company-reported)
- 30.9%
- OSWorld 2.0 score, Opus 5.5 (as listed by H)
- 81.8%
- Training tasks generated by H's Agentic Task Factory (company figure)
- About 10,000
Why CEOs should care
For COOs and CIOs with old software that has no API, computer-use models are the practical route to automation, because the agent works the same screens your staff do. An open-weight option matters here: you can run it on your own hardware, keep screenshots of internal systems inside your network, and avoid sending sensitive screens to an outside provider. Ask your automation team which workflows are blocked today only because the software cannot be called by code.
For CFOs, the cost claims need checking before any budget moves. H says Holo4 competes at a much lower cost per task, but The Register read the company's own charts as showing higher costs per task than some rivals in many cases, and said this was likely because the model uses more reasoning tokens. Run a pilot on your own workflows and measure cost per completed task, not per token or per model call.
For CISOs, an agent that can click through any interface is also an agent that can click through the wrong one. Before deployment, decide which accounts the agent uses, what it is allowed to touch, how every action is logged, and who reviews failures. The same controls apply whether the model runs in your data center or through H's API.
The bigger picture
H is one of several developers working on computer use. The Register noted that Amazon Web Services announced its own computer-use models at its re:Invent conference last year, and that OpenAI, Google and Anthropic are also investing in the capability. Holo4's pitch is that a smaller model can do much of the same work. The Register estimated that a 24GB Nvidia RTX 3090 graphics card should be able to run the models at 4-bit precision, which would put on-premises testing within reach of many IT teams.
What’s next
H said it will release optimized DSpark drafter checkpoints in the coming days, which it expects to speed up inference; The Register explained this uses a technique called speculative decoding. The company also said it will report Holo4's AutomationBench results on that benchmark's private test set once evaluated. The license terms for the weights were not stated in the sources we reviewed, so buyers should confirm them before commercial use.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
