H Company's Holo4-27B Scores 61.7% on OSWorld 2.0, and Its Model Card Bars Commercial Use of an Apache 2.0 Base
A.I. / news
H Company's Holo4-27B Scores 61.7% on OSWorld 2.0, and Its Model Card Bars Commercial Use of an Apache 2.0 Base
The September 28 release fine-tunes Alibaba's Qwen3.8-27B for computer use. The base allows commercial work; the derivative is CC BY-NC 4.0.
H Company released Holo4, a family of computer-use models, on September 28, with the 27-billion-parameter Holo4-27B scoring 61.7% on OSWorld 2.0 at $1.22 a task, according to the company's Hugging Face post. The 27B's model card lists the licence as CC BY-NC 4.0, which is non-commercial.
The card names its base model as Qwen3.8-27B. That model's own card lists Apache 2.0. A company can fine-tune an Apache model and release the result under stricter terms, but the reader who assumes the Qwen licence carries over will be wrong.
What shipped on September 28
The family has three models, all with open weights on Hugging Face in BF16, FP8, NVFP4 and 4-bit GGUF formats, per the post.
| Model | Architecture | OSWorld 2.0 |
|---|---|---|
| Holo4-27B | 27B dense | 61.7% |
| Holo4-35B-A3B | 35B mixture of experts | 30.9% |
| Holotron4-30B-A3B | 33B listed | Not given |
The card for the 27B gives a 262,144-token context window and an AutomationBench score of 45.4% at $0.05 a task. The post says the models act through "GUIs, code, MCP and APIs." The Terminal could not confirm the licences of the other two models from the pages it read, and the post as retrieved does not state them.
How the score compares
The post sets Holo4 against Claude Opus 5.5, which it lists at 81.8% on OSWorld 2.0. All of these numbers are H Company's own, and the pages read for this story cite no outside party running them.
- Claude Opus 5.581.8 %
- Holo4-27B61.7 %
- Holo4-35B-A3B30.9 %
Source: H Company post on Hugging Face, accessed 2026-10-05
The 27B sits 20.1 points below the closed model on that benchmark. H Company's argument is cost: the post says Holo4 reaches this level with "orders of magnitude fewer parameters" than frontier systems. The post gives the 27B's per-task cost, $1.22, but the figure for the Claude comparison was not in the text retrieved.
Training and the licence question
The post describes supervised fine-tuning on 127 billion tokens, then reinforcement learning from an internal Agentic Task Factory that produced about 10,000 tasks across web applications, desktop environments and MCP servers. It credits more than 77 contributors from H Company.
The card says Holo4-27B shows "significant improvements over their Qwen base models." Beyond the licence line, it does not list the training data.
CC BY-NC 4.0 permits copying and adaptation with attribution and bars commercial use. A team wanting to run Holo4-27B inside a paid product would need terms from H Company, which sells access through its H Models API. A team wanting the same size class without that condition can start from the Qwen base, as Cloudflare's Clef did under Apache 2.0. Whether weights are downloadable is a separate question from what the licence allows, as the site's piece on Runway's Praxis-1 showed.
H Company has not said whether the 35B and Holotron models carry the same non-commercial licence. It has not published a date for any updated release.
Sources
More in A.I.
- 01Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX SparkThe model card lists 3.4B active parameters, a 1M-token window and a 20GB checkpoint, but does not say who ran most of its benchmark rows.
- 02Anthropic Closed Its Debut at $97.17, 17% Over the $83 IPO Price, as Nasdaq Fell 2%The $81.34 billion offering opened 34% higher, then gave back most of the gain while Nvidia, AMD and Micron dropped.
- 03Google's Gemini Tool Retrieval Preview Adds 26 Tokens Per Idle Tool and Publishes No Savings FigureAnthropic and OpenAI each reported token reductions for the same defer_loading flag. Google's November 19 post reports none.
- 04OpenAI Says New Safeguards Cut Reasoning Extraction From 54.0% to 0.2%The figures are from OpenAI's own red-team evaluation, and the post names DeepSeek and Moonshot AI as distillers of U.S. models.