Original research brief
Holo4 Series Introduces Generalist Agentic AI Models for Computer Use
A report on the Holo4 series of agentic AI models, which are designed to interact with software through multiple interfaces like GUIs, code, and APIs for…
What this piece is grounded in
01
Holo4 comes in two sizes: 27B dense and 35B-A3B Mixture of Experts. (Hugging Face)
02
Holo4 interacts with software through any available interface: GUIs, code, MCP and APIs. (Hugging Face)
03
On OSWorld 2.0, Holo4 27B scores 61.7% against 81.8% for Opus 5.5. (Hugging Face)
04
Holo4 runs on desktops, on the web, on Android, in a code sandbox and against business APIs. (Hugging Face)
01 / FIELD NOTE
What Changed
The source company announced Holo4, a new series of agentic models. According to the Hugging Face source, Holo4 comes in two sizes: 27B dense and 35B-A3B Mixture of Experts. The models are built to interact with software through any available interface: GUIs, code, MCP and APIs. The source states Holo4 runs on desktops, on the web, on Android, in a code sandbox and against business APIs. It is the same model in each case. The company is also releasing an updated version of Holotron 3, called Holotron4 Nano.
02 / FIELD NOTE
Why It Matters for Practical AI Work
Editorial interpretation: The ability to use a single model across different interfaces, as described in the source, could reduce the engineering effort of selecting and integrating specialized models for GUI, code, or API tasks. This is tied to the source fact that a single business task can require combining these different approaches. The source's claim that Holo4 models improve significantly over their Qwen base suggests potential for more capable automation from a known model family.
03 / FIELD NOTE
What a Reader Can Verify
The source provides specific benchmark scores for verification. On OSWorld 2.0, Holo4 27B scores 61.7% against 81.8% for Opus 5.5, and Holo4 35B-A3B reaches 30.9%. The company states it open-sources every trajectory behind its scores on public benchmarks. Readers can replay each step at trajectories.hcompany.ai or download them from Hugging Face. The source also includes detailed examples of the model performing tasks like 3D modeling in FreeCAD and game design in Godot, with call and token counts provided.
04 / FIELD NOTE
What Remains Unclear
The supplied feed entry is a concise summary; consult the canonical article for full context. The source notes that Holo4 will be reported on the AutomationBench private set once it is evaluated, indicating a pending result. The exact pricing for the H Models API, beyond the statement that Holo4 is priced at H Models API rates, is not detailed in the provided text. The comparative cost claims reference internal harness measurements and list prices from other providers, the specifics of which require the full source.
Questions readers ask
What are the main interfaces Holo4 can work with?
According to the source from Hugging Face, Holo4 clicks and types on a screen, writes and runs its own code, and calls MCP or API tools. It is designed to use whichever interface fits the task, including GUIs, code, MCP and APIs.
How does Holo4's performance compare to other models?
The source reports that on the OSWorld 2.0 benchmark, Holo4 27B scores 61.7% against 81.8% for Opus 5.5, and Holo4 35B-A3B reaches 30.9%. The company states Holo4 trails only the strongest closed models on long workflows but does so with orders of magnitude fewer parameters and at a much lower cost.