Holo4 uses one model across screens, code and APIs, but its two weights have different trade-offs

A business workflow can move from a browser form to a terminal and then to an API. H Company’s Holo4 launch argues that an agent should not need a different model for each interface. Its new computer-use family has a 27-billion-parameter dense model and a 35-billion-parameter mixture-of-experts (MoE) model with 3 billion active parameters. Both are available through the H Models API and as downloadable weights. The practical choice is not merely speed versus quality: the two checkpoints have different licenses, and the cheaper model falls much further behind on H Company’s long-workflow test than its basic desktop score suggests.

One model can choose the interface

Holo4 is trained to click and type in GUIs, write and run code, and call MCP or API tools within a task. H says it works across desktop, web, Android and code sandboxes without changing models for each environment. Its Models API documentation describes an OpenAI-compatible call pattern in which the agent receives tools and emits tool calls; the surrounding harness still has to provide screenshots, execute actions and manage the conversation. This is a model-and-agent stack, not a promise that downloading a checkpoint alone will control a desktop.

The two model cards list a maximum context configuration of 262,144 tokens for the 27B model and 262,144 for the 35B-A3B model. H’s quickstart uses the model ID holo4-27b with the OpenAI client pointed at H’s API endpoint. It says both Holo4 models require the paid tier; its free-tier example uses the older holo3-1-35b-a3b. The launch also makes BF16, FP8, NVFP4 and 4-bit GGUF variants available on Hugging Face.

H describes a training route that starts with supervised trajectories across interfaces and then trains separate reinforcement-learning specialists for desktop/web and terminal/MCP/API work before merging them into one model. It also credits harness changes, including memory across hundreds of steps and a shell on the desktop machine, for improved task performance. The reported results therefore reflect the released model with its harness, not a model-only desktop-control measurement.

The long-workflow gap is larger than the headline desktop gap

On H Company’s OSWorld comparison, Holo4 27B scores 85.2% and 35B-A3B scores 80.8%. On its OSWorld 2.0 long-workflow measure, however, their partial scores are 61.7% and 30.9%, with success rates of 41.5% and 12.3%, respectively. H lists costs of $1.22 and $0.61 per task on that test at its API rates. A lower-priced run is not the same thing as comparable completion on longer jobs.

These are H Company’s reported evaluations, not AgentRiot tests. The launch’s methodology notes say Holo4 scores in its harness are usually means over two to four runs, but OSWorld 2.0 and ALE-CLI use a single run. Its frontier reference scores come from public sources using differing harnesses, effort settings and in some cases task subsets. Its 600-task AutomationBench public set also includes 480 tasks in the split from which H collected training data; the company separately reports a held-out 120-task score. The useful comparison within the release is the marked difference between Holo4’s two models under H’s own conditions, not a universal frontier ranking.

Open weights, different permissions

The 27B model card lists CC BY-NC 4.0 for Holo4’s weights: noncommercial use. The 35B-A3B card lists Apache 2.0. A team planning commercial self-hosting should not treat the shared “open weights” label as a common license. Both cards identify Qwen base architectures, but the Holo4 weight licenses are the relevant terms for these releases.

For hosted use, the launch page lists 27B at $0.40 per million input tokens and $3.00 per million output tokens, with cached input at $0.04. It lists 35B-A3B at $0.30 input and $2.00 output, with cached input at $0.03, on the same per-million basis. Those are published API list rates, not a guarantee about the cost of a completed workflow; the company’s own OSWorld 2.0 results illustrate how success and token use alter that calculation.

H has also released trajectories behind its public-benchmark scores for inspection. That is a useful transparency step, though it does not make mixed-harness competitor numbers directly comparable. For a builder, Holo4’s immediate question is concrete: does the same agent need to navigate a screen and call tools, and is the 27B model’s stronger long-run result worth its cost and noncommercial weight license? The published materials supply a starting point; they do not replace testing on the workflow that actually matters.