Holo4 opens the agent workbench through weights, traces and harnesses

A computer-use model is only one component of an agent. It must receive screenshots or structured tool results, decide what to do, send clicks, keystrokes, code or API calls, and inspect what happened next. H Company’s Holo4 release is notable because it publishes several layers of that system: model weights, a harness, benchmark results and thousands of recorded runs.[1][2][4]

That makes it possible to ask a better question than whether the weights are “open.” Can a team inspect how the agent acted, reproduce the environment, understand the licence, change the runtime and test failures that matter to its own workflow? Holo4 provides more material for that investigation than a checkpoint alone.

Openness has several layers

The licence difference is operationally important. The Holo4-27B model card marks the weights CC BY-NC 4.0, which restricts commercial use. The inspected 35B-A3B FP8 card uses Apache 2.0. A company choosing a model for production therefore cannot treat the two checkpoints as interchangeable merely because they belong to one family.[2][3]

Trajectories are more valuable than a single score

H reports 85.2% on OSWorld for Holo4 27B at $0.08 per task. On OSWorld 2.0, it reports 61.7% at $1.22 per task, while the 35B-A3B model reports 30.9% at $0.61. On AutomationBench, the corresponding results are 45.4% at $0.05 and 34.5% at $0.02. These vendor-reported figures show why a model can occupy different quality–cost positions depending on the benchmark.[1][2][3]

A trajectory reveals information the final success rate cannot: whether the agent wandered, touched unrelated controls, retried the same action, relied on a fragile visual cue or succeeded only after an unsafe intermediate step. The released dataset makes those questions inspectable for the published runs. It does not establish that every task was representative or that a different application will behave the same way.[4]

What a serious pilot should replay

Choose five workflows that reflect the intended deployment and define forbidden actions before running the model. Capture the full observation–action sequence, not just the final screen. Then score task completion, unnecessary calls, intervention time, policy violations and recovery after a changed interface. Where possible, replay the same tasks with both models and a fixed harness so the comparison does not quietly mix model quality with runtime differences.

Holo4’s strongest contribution may be this broader release surface. Weights let builders run the model; the harness shows how it is used; trajectories make published behavior inspectable; and the licences force a concrete deployment decision. That is a more demanding—and more useful—definition of an open agent stack.

Report an error or suggest a correction