Hugging Face has released Holo4, a pair of agentic models designed to move between graphical interfaces, code, MCP and APIs within the same workflow. The useful shift is flexibility: one model can tackle a business task across different software surfaces, rather than being tied to a screen or a tool call.
Hugging Face Watch analysis
What happened
Holo4 comes in 27B dense and 35B-A3B Mixture of Experts versions. Hugging Face says both are available through its H Models API, with weights in several formats, including FP16, FP8 and GGUF. It also released Holotron4 Nano, an updated agent model based on Nemotron 3 Nano Omni.
The company says Holo4 can work on desktops, the web, Android, code sandboxes and business APIs. Its published benchmark figures are more mixed than the broad pitch: on OSWorld 2.0, Hugging Face reports 61.7% for Holo4 27B and 30.9% for Holo4 35B-A3B, compared with 81.8% for Opus 5.5. The company says the models cost much less per task, but notes that the comparisons draw on different releases, harnesses and task subsets.
Why it matters
The ambition is a generalist agent that can switch tools as a task demands, rather than needing a separate model for every interface. That could make agents easier to fit into real business workflows, where work tends to wander between apps with little regard for neat product categories.
Hugging Face also says it is publishing the trajectories behind its benchmark scores, so readers can inspect the steps behind the results. The figures remain company-reported comparisons, not a clean independent head-to-head; task costs and performance depend on the harness and evaluation set.
Our read
This is a substantive release, not just another agent with a fresh badge: it brings multiple interaction modes together and offers model weights for local use. The most interesting promise is practical breadth. The score to keep in perspective is the one that makes the model sound closest to the frontier: Hugging Face itself reports a sizeable OSWorld gap to Opus 5.5.
If you are choosing an agent for real work, look beyond the headline score. Try the workflow you actually need, and check the published trajectories and costs against the same tasks and setup.
What to watch
- Whether independent evaluations reproduce the reported performance and cost comparisons.
- How well Holo4 handles long workflows that switch between interfaces.
- Whether open trajectory data makes the results easier to inspect and reproduce.
Discussion spark: For business agents, is broad competence across screens and APIs more valuable than a higher score on a narrower benchmark?
Sources and evidence
- Holo4: powering generalist computer-use agents (28 September 2026, 09:44 UTC)
not affiliated with or endorsed by Hugging Face