H Company has released Holo4, a family of AI models it says can work through screens, write code and use direct connections to other software. Its September 28 announcement includes the 27B and 35B-A3B models, with downloadable weights and API availability.
The useful idea is that the agent can choose how to do a task. Some software gives it a direct way to request information or make a change. Elsewhere, it may need to use the visible screen. H Company says Holo4 can move among those approaches, including MCP connections, which let AI tools interact with other software.
Picture a job that involves reading information in a browser, calculating something and entering the result into another application. That’s an example of the kind of mixed work this direction could help with, not a task I’ve tested on Holo4. An agent that can follow the work across those steps could save a person from moving each result by hand.
H Company has shared benchmark results and records showing the steps its agents took. Those records give others something to inspect. The performance claims are still the company’s, and a benchmark doesn’t establish how reliably the model will handle someone’s actual workflow.
There’s also a licensing detail worth catching before trying the downloadable models for business use. The 27B model is listed under CC BY-NC 4.0, which restricts commercial use. The 35B-A3B model is listed under Apache 2.0. They shouldn’t be treated as having the same terms just because both can be downloaded.
I find this direction impressive. The question I’d want to test is whether the agent can carry a task accurately from one application to the next, with the right permissions and a clear way for a person to step in. I haven’t run that test. That’s what would tell me more than its place on a leaderboard.