H Company Releases Holo4 Computer-Use Models
The new open models can move between graphical interfaces, code, MCP and APIs, with weights and evaluation trajectories available for inspection.
Edited by Tyronne Panaino
H Company released Holo4 on September 28, 2026, as a family of agentic models designed to work across graphical interfaces, code, Model Context Protocol tools and conventional APIs. The release includes a 27-billion-parameter dense model, a 35B-A3B mixture-of-experts model and an updated Holotron4 Nano variant.
Both main Holo4 models are available through the H Models API, while model collections and execution trajectories are also available for inspection. That combination makes the release relevant to developers comparing hosted computer-use systems with models they can examine and run through their own agent harnesses.
One model across several interfaces
H Company's central design claim is that a single Holo4 model can choose among clicking and typing on a screen, writing and running code, or calling MCP and API tools. The company says the same model can operate on desktops, the web, Android, in a code sandbox and against business APIs.
That is a meaningful architectural target because real workflows rarely stay inside one interface. An agent might need to read a web application with no API, transform data in code and then call an authenticated business tool. A system trained only for screen control or only for tool calls can stall when the next step falls outside its preferred interface. Holo4 is intended to reduce that handoff problem.
The models were post-trained with supervised and reinforcement learning on a collection of environments and tasks, including work generated through H Company's Agentic Task Factory. The release also includes an updated Nano model produced by applying the post-training stack to NVIDIA's Nemotron 3 Nano Omni base.
Open trajectories improve inspectability
H Company links model variants, a full collection of weights, a trajectory viewer and a downloadable trajectory dataset. Publishing trajectories matters because a benchmark result alone does not show whether an agent reached an answer through stable, safe steps or through brittle shortcuts. Inspectable action traces give researchers and engineering teams more material for reproducing failures and testing their own policies.
Open trajectories do not make the reported results independent. H Company produced the models, selected the harness and published the comparisons. The release itself notes that model releases, harnesses and task subsets differ across its cost-and-performance charts. It also says Holo4 has not yet been evaluated on the private AutomationBench set, so results from the public set should not be treated as a substitute.
What developers still need to verify
Internal confidence is medium because one first-party technical article establishes the release and resources but no independent evaluation was fetched in this run. Teams evaluating Holo4 should test recovery from failed actions, permission boundaries, long-task consistency and behaviour when a workflow switches between GUI and authenticated tools.
The next useful checkpoints are reproducible third-party tests using the published trajectories, results on private or contamination-resistant task sets, and deployment documentation covering access control and observability. Those checks will show whether the cross-interface design translates into dependable production work rather than isolated demonstrations.
Status
Confirmed. H Company has released Holo4 model variants, an API path and inspection resources. Capability and cost comparisons remain developer-reported and are constrained by differing evaluation setups.
Sources
Update note: Last reviewed 2026-10-04. We will revise this post when independent evaluations or additional production deployment evidence become available.
Sources
- H Company — Holo4 release — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.