H Company released Holo4 on September 28, 2026: open-weight models for computer-use agents in two sizes, a 27-billion-parameter dense model and a mixture-of-experts model named Holo4-35B-A3B. It also released Holotron4 Nano, a smaller agent model built on NVIDIA's Nemotron 3 Nano Omni. The weights are on Hugging Face, but the three models carry three different licenses, and one of them does not allow commercial use.
In its announcement on Hugging Face, H Company describes Holo4 as a series of agentic models that operate software through whatever interface is available: clicking and typing on a screen, writing and running code, or calling MCP and API tools. The pitch is that one model handles desktops, the web, Android, code sandboxes and business APIs, so an agent does not need a separate model for each surface.
Both Holo4 models are fine-tuned from Alibaba's Qwen models. The Holo4-35B-A3B model card says it is built on Qwen3.6-35B-A3B, and the announcement shows Holo4-27B alongside Qwen3.8 27B as its base. H Company says it trained them with supervised and reinforcement learning on a large set of environments, including about 10,000 tasks produced by its own task-generation system, and rebuilt its agent harness to keep track of hundreds of steps.
The model cards list a context length of 262,144 tokens, and image and text input with text output.
| Model | Architecture | License (per Hugging Face model card) |
|---|---|---|
| Holo4-27B | 27B dense | CC BY-NC 4.0 (non-commercial) |
| Holo4-35B-A3B | Mixture of experts (35B-A3B) | Apache 2.0 |
| Holotron4 Nano (repository Holotron4-30B-A3B) | Built on NVIDIA Nemotron 3 Nano Omni | NVIDIA Open Model Agreement |
Licenses read from each model card on October 10, 2026.
This is the most important practical detail of the release. Holo4-27B, the stronger of the two on OSWorld 2.0 in H Company's results, is licensed for non-commercial use only. Holo4-35B-A3B uses Apache 2.0, a permissive open-source license that allows commercial use. Holotron4 Nano uses NVIDIA's own model agreement, which has its own terms to review. A team that tests the 27B model and then plans to ship it in a commercial product needs either a different license arrangement or the H Models API.
Weights are published in BF16, FP8, NVFP4 and 4-bit GGUF formats, according to the announcement. Those are packaging variants of the same models, not separate releases. Both Holo4 sizes are also available through H Company's hosted H Models API; the announcement does not list API prices.
On OSWorld 2.0, a computer-use benchmark, H Company reports 61.7% for Holo4-27B and 30.9% for Holo4-35B-A3B, against 81.8% for Claude Opus 5.5. It describes Holo4 as trailing only the strongest closed models on long workflows, at a much lower cost per task. The gap between the two Holo4 sizes is large: the 35B-A3B mixture-of-experts model scores about half as much as the dense 27B model on that benchmark.
H Company is unusually transparent about method. It publishes every trajectory behind its public benchmark scores, viewable step by step or downloadable as a dataset, and its chart notes explain which competitor numbers come from launch data, leaderboards or its own harness. It also notes that releases, harnesses and task subsets differ across those sources. These remain H Company's measurements, and this article does not reproduce them.
Holotron4 Nano is described as a demonstration that H Company's post-training recipe transfers to other base models. The announcement reports gains over Nemotron 3 Nano Omni on GUI, MCP, API and coding-sandbox environments.
Teams building computer-use or browser agents that want to self-host have a strong reason to test Holo4. Few open-weight models target GUI, code and API interfaces together.
Commercial products should start from Holo4-35B-A3B (Apache 2.0), or review the NVIDIA agreement for Holotron4 Nano, and treat Holo4-27B as a research and evaluation model unless H Company grants other terms.
Teams comparing against closed models can weigh Holo4 against closed computer-use options such as Claude Haiku 5.5, which Anthropic's announcement describes as working especially well for browser use, and Claude Opus 5.5, the reference point in H Company's own chart.
When was Holo4 released?
September 28, 2026.
Is Holo4 open source?
The weights are public, but the licenses differ: Holo4-35B-A3B is Apache 2.0, Holo4-27B is CC BY-NC 4.0 (non-commercial), and Holotron4 Nano uses the NVIDIA Open Model Agreement.
What is Holotron4 Nano?
A smaller agent model from H Company, built on NVIDIA Nemotron 3 Nano Omni and published as Holotron4-30B-A3B on Hugging Face.
Discover more content:
