When the task changes interface
H Company says Holo4 can move among clicking on a screen, running code, and calling MCP or API tools. For a builder, that shifts attention to which access path a task needs at each step. An API can reach a service without a screen; a graphical interface can reach software that exposes no suitable tool. The same task can cross both boundaries.[1]
The company has published weights for two models and opened API access. The smaller dense model has 27 billion parameters. H Company describes thousands of tasks generated in its training setup, plus changes to memory and shell access for long runs. Those pieces matter alongside model weights: diagnosing a failed step requires seeing which interface the agent chose and what context it retained.[1]
Traces offer a way to inspect the claim
H Company says the execution traces for its public benchmark tasks can be downloaded and replayed. Builders can inspect the steps behind a score, including how an agent moves between interfaces. The company also notes that models in its comparison chart come from different task sets, releases and execution setups. A single score therefore cannot isolate the benefit of cross-interface switching.[1]
The useful test for Holo4 is to run the same task first through one interface, then with screen and tool switching enabled. Counting completed tasks alongside reversals, human review and compute use would reveal what the new layer contributes. H Company’s open traces provide material for that inspection; the announcement does not include an independent matched result.[1]