Holo4-27B

Holo4-27B

Looks at screenshots and decides the next click, keystroke, code run or tool call, built on Qwen3.8-27B. A computer-use model for web, desktop and mobile agents, under a non-commercial license.

Made by Hcompany on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Holo4-27B is a vision-language model for computer use, built by H Company on Qwen3.8-27B. It reads screenshots and tool results, then asks for clicks, typing, code and tool calls. A separate agent package carries out those requests. The card reports 85.2% on OSWorld at $0.08 per task.

Who it suits: It suits developers and researchers who build agents that operate real software, on the web, on a desktop or on a phone. The license is non-commercial.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Software | Agent developers | Drive a graphical interface from screenshots | An agent that fills in a web form
Research | Computer-use researchers | Test a 27B model on desktop tasks | Run OSWorld-style tasks
Business automation | Workflow teams | Try agents on multi-step business workflows | A task across a browser and a tool server

If you need a model for a commercial product, you can skip it.

Where to use it:

• The hai-agents Python package, which the card names as the way to run the model
• The transformers library, on Windows, macOS or Linux with Python
• The H Models API, a hosted option that the card links
• Web, desktop and mobile apps, the targets the card lists

Title: Jobs it does
Use case | Who | What they get back
Operating a graphical interface | Agent builders | Clicks and typing planned from a screenshot
Writing and running code | Developers | Code the model requests inside a task
Calling tools | Workflow teams | Tool calls for web, desktop and MCP workflows

How to set it up:

• Python on Windows, macOS or Linux
• A GPU with room for the 54.7 GB BF16 weights, or one of the smaller FP8, NVFP4 or Q4 GGUF editions
• The hai-agents package, which sends screenshots and runs the model's requests

1. Install the Hugging Face Hub tools.
2. Download the model files.
3. Connect the model to the hai-agents package, following the model card.

pip install -U huggingface_hub
hf download Hcompany/Holo4-27B

How to use it: The agent package sends a screenshot and tool results to the model. The model answers with a click, some typing, code or a tool call, and the package runs it. Then the next screenshot goes back. The cut card text shows no further commands.

Title: What the model can ask for
Tool | What it does
Graphical interface | Clicks and typing on what the screenshot shows
Code | Runs code that the model requests
Tool calls | Calls tools, including MCP tools in business workflows

hf download Hcompany/Holo4-27B

Example: You ask the agent to model the Eiffel Tower in FreeCAD, with legs, platforms and a mast. The model works through the program by clicking, typing and running code until a tower model exists. The card shows this in a video.

For newcomers and experts: If you are new, this is an AI that looks at your screen and decides what to click or type next. The first small thing to try is reading the card's model summary before you download the 54.7 GB files.

If you know the field, it is a 27B dense Qwen3.8 model with a configured context of 262,144 tokens. The card compares it with its Qwen base on OSWorld and on Agentic Task Factory, a held-out set of business workflows. The trade-off is evidence: the benchmark table is an image, and the license bars commercial use.

Screenshot and tool results go to the model
The model asks for a click, typing or code
The agent package runs the request
A new screenshot goes back

Good to know:

• Family: Holo4 also has a 35B-A3B mixture-of-experts model and Holotron4-30B-A3B, each with FP8 and other editions.
• Editions of this model: BF16, FP8, NVFP4 and a Q4 GGUF.
• Reported cost: 85.2% on OSWorld at $0.08 per task.
• License: cc-by-nc-4.0, so it is non-commercial.
• Cost to run: a large GPU, or the hosted H Models API, whose price the card does not give.
• Data it reaches: it sees the screenshots you send and can click, type and run code, so what it reaches depends on the environment you give it.
• Not for: commercial use, which the license bars.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.4B
• Weights: about 54.7 GB in BF16; about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,486
• Likes: 105
• License: cc-by-nc-4.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 14
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, multimodal, computer-use, agent, conversational

Links

• Model: https://huggingface.co/Hcompany/Holo4-27B
• Files: https://huggingface.co/Hcompany/Holo4-27B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Hcompany/Holo4-27B

Then follow the usage in its model card. Nothing in this kit runs by itself.