llm.create_chat_completion(
messages = [
{
"role": "user",
"content": [
{
"type": "text",
"text": "Describe this image in one sentence."
},
{
"type": "image_url",
"image_url": {
"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
}
}
]
}
]
)Holo2-30B-A3B GGUF Models
Model Generation Details
This model was generated using llama.cpp at commit 7d77f0732.
Quantization Beyond the IMatrix
I've been experimenting with a new quantization approach that selectively elevates the precision of key layers beyond what the default IMatrix configuration provides.
In my testing, standard IMatrix quantization underperforms at lower bit depths, especially with Mixture of Experts (MoE) models. To address this, I'm using the --tensor-type option in llama.cpp to manually "bump" important layers to higher precision. You can see the implementation here:
👉 Layer bumping with llama.cpp
While this does increase model file size, it significantly improves precision for a given quantization level.
I'd love your feedback—have you tried this? How does it perform for you?
Click here to get info on choosing the right GGUF model format
Holo2: Foundational Models for Navigation and Computer Use Agents
Model Description
Holo2 represents the next major step in developing large-scale Vision-Language Models (VLMs) for multi-domain GUI Agents. These agents can operate real digital environments specifically web, desktop, and mobile by interpreting interfaces, reasoning over content, and executing actions.
Our Holo2 family emphasizes navigation and task execution across diverse real and simulated environments, extending beyond static perception to multi-step, goal-directed behavior.
It builds upon the strengths of Holo1.5 in UI localization and screen content understanding, with major improvements in policy learning, action grounding, and cross-environment generalization.
The Holo2 series comes in three model sizes:
- Holo2-4B: fully open under Apache 2.0
- Holo2-8B: fully open under Apache 2.0
- Holo2-30B-A3B: research-only license (non-commercial). For commercial use, please contact us.
These models are designed to provide reliable, accurate, and efficient foundations for next-generation CU agents, like Surfer-H.
- Developed by: H Company
- Model type: Vision-Language Model for Navigation and Computer Use Agents
- Fine-tuned from model: Qwen/Qwen3-VL-30B-A3B-Thinking
- Blog Post: https://www.hcompany.ai/blog/holo2
- License: Apache 2.0 License
Get Started with the Model
Please have a look at the cookbook in our repo where we provide examples for both self-hosting and API use!
Training Strategy
Our models are trained using high-quality proprietary data for UI understanding and action prediction, following a multi-stage training pipeline. The training dataset is a carefully curated mix of open-source datasets, large-scale synthetic data, and human-annotated samples. Training proceeds in two stages: large-scale supervised fine-tuning, followed by online reinforcement learning (GRPO) yielding SOTA performance in interpreting UIs and performing actions on large, complex screens
Results
Holo2: Navigation Performance
Navigation evaluates an agent’s ability to complete real or simulated tasks through multi-step reasoning and action.
Holo2 models show significant improvements in navigation efficiency and task completion rates, particularly in unseen and complex environments.
Benchmarks include WebVoyager, WebArena, OSWorld, and AndroidWorld, testing the models’ abilities across web, operating system, and mobile platforms.
| Model | WebVoyager | WebArena | OSWorld | AndroidWorld | Average |
|---|---|---|---|---|---|
| Holo2-30B-A3B | 83.0% | 46.3% | 37.4% | 71.6% | 59.6% |
| Holo2-8B | 80.2% | 42.2% | 39.9% | 60.4% | 55.7% |
| Holo2-4B | 80.2% | 41.0% | 37.7% | 64.6% | 55.9% |
| Holo1.5-7B | 65.9% | 23.4% | 6.4% | 32.7% | 32.1% |
| Holo1.5-3B | 56.1% | 15.4% | 5.8% | 27.5% | 26.2% |
| Qwen3-VL-30B-A3B-Thinking | 76.1% | 45.0% | 36.6% | 62.9% | 55.1% |
| Qwen3-VL-8B-Thinking | 72.0% | 31.9% | 28.8% | 52.6% | 46.3% |
| Qwen3-VL-4B-Thinking | 67.5% | 31.5% | 24.1% | 45.7% | 42.2% |
Table 1: Navigation benchmark scores. Bold values will denote state-of-the-art once final evaluations are available.
All external model scores are reproduced internally in the Surfer 2 agent, to allow for fair comparison
Holo2: SOTA UI Localization
UI Localization measures how precisely an agent can locate on-screen elements—buttons, inputs, links—necessary for accurate interaction.
Holo2 continues to set new standards for localization accuracy across web, OS, and mobile benchmarks.
| ScreenSpot-Pro | OSWorld-G | Showdown | Ground-UI-1K | WebClick-v1 | ScreenSpot-v2 | Average | |
|---|---|---|---|---|---|---|---|
| Holo2-30B-A3B | 66.1% | 76.1% | 77.6% | 85.4% | 91.3% | 94.9% | 81.90 |
| Holo2-8B | 58.9% | 70.1% | 72.5% | 83.8% | 89.5% | 93.2% | 78.00 |
| Holo2-4B | 57.2% | 69.4% | 74.7% | 83.3% | 88.8% | 93.2% | 77.77 |
| Holo1.5-72B | 63.3% | 71.8% | 76.8% | 84.5% | 92.4% | 94.4% | 80.52 |
| Holo1.5-7B | 57.9% | 66.2% | 72.1% | 84.0% | 90.2% | 93.3% | 77.28 |
| Holo1.5-3B | 51.4% | 61.5% | 67.5% | 83.2% | 81.4% | 91.6% | 72.77 |
| Qwen3-VL-30B-A3B-Thinking | 49.9% | 65.8% | 71.2% | 84.2% | 89.5% | 91.8% | 75.40 |
| Qwen3-VL-8B-Thinking | 38.5% | 56.0% | 64.2% | 83.6% | 85.9% | 91.5% | 69.95 |
| Qwen3-VL-4B-Thinking | 41.4% | 56.4% | 66.6% | 84.1% | 85.8% | 90.0% | 70.72 |
| Qwen2.5-VL-72B | 55.6% | 62.0% | 41.0% | 85.4% | 88.3% | 93.3% | 70.93 |
| Qwen2.5-VL-7B | 29.0% | 40.6% | 52.0% | 80.7% | 76.5% | 85.6% | 60.73 |
| Qwen2.5-VL-3B | 29.3% | 34.3% | 50.3% | 76.4% | 71.2% | 80.7% | 57.03 |
| UI-TARS-1.5-7B | 39.0% | 61.0% | 58.0% | 84.0% | 86.1% | 94.0% | 70.35 |
| UI-Venus-72B | 61.9% | 70.4% | 75.6% | 75.5% | 77.0% | 95.3% | 75.95 |
| UI-Venus-7B | 50.8% | 58.8% | 67.3% | 82.3% | 84.4% | 94.1% | 72.95 |
Table 2: Localization benchmark scores for leading models.
Accuracy of our and competitors' models on UI Localization benchmarks.
Citation
@misc{hai2025holo2modelfamily,
title={Holo2 - Open Foundation Models for Navigation and Computer Use Agents},
author={H Company},
year={2025},
url=https://huggingface.co/collections/Hcompany/holo2,
}
<!--End Original Model Card-->
---
# <span id="testllm" style="color: #7F7FFF;">🚀 If you find these models useful</span>
Help me test my **AI-Powered Quantum Network Monitor Assistant** with **quantum-ready security checks**:
👉 [Quantum Network Monitor](https://readyforquantum.com/?assistant=open&utm_source=huggingface&utm_medium=referral&utm_campaign=huggingface_repo_readme)
The full Open Source Code for the Quantum Network Monitor Service available at my github repos ( repos with NetworkMonitor in the name) : [Source Code Quantum Network Monitor](https://github.com/Mungert69). You will also find the code I use to quantize the models if you want to do it yourself [GGUFModelBuilder](https://github.com/Mungert69/GGUFModelBuilder)
💬 **How to test**:
Choose an **AI assistant type**:
- `TurboLLM` (GPT-4.1-mini)
- `HugLLM` (Hugginface Open-source models)
- `TestLLM` (Experimental CPU-only)
### **What I’m Testing**
I’m pushing the limits of **small open-source models for AI network monitoring**, specifically:
- **Function calling** against live network services
- **How small can a model go** while still handling:
- Automated **Nmap security scans**
- **Quantum-readiness checks**
- **Network Monitoring tasks**
🟡 **TestLLM** – Current experimental model (llama.cpp on 2 CPU threads on huggingface docker space):
- ✅ **Zero-configuration setup**
- ⏳ 30s load time (slow inference but **no API costs**) . No token limited as the cost is low.
- 🔧 **Help wanted!** If you’re into **edge-device AI**, let’s collaborate!
### **Other Assistants**
🟢 **TurboLLM** – Uses **gpt-4.1-mini** :
- **It performs very well but unfortunatly OpenAI charges per token. For this reason tokens usage is limited.
- **Create custom cmd processors to run .net code on Quantum Network Monitor Agents**
- **Real-time network diagnostics and monitoring**
- **Security Audits**
- **Penetration testing** (Nmap/Metasploit)
🔵 **HugLLM** – Latest Open-source models:
- 🌐 Runs on Hugging Face Inference API. Performs pretty well using the lastest models hosted on Novita.
### 💡 **Example commands you could test**:
1. `"Give me info on my websites SSL certificate"`
2. `"Check if my server is using quantum safe encyption for communication"`
3. `"Run a comprehensive security audit on my server"`
4. '"Create a cmd processor to .. (what ever you want)" Note you need to install a [Quantum Network Monitor Agent](https://readyforquantum.com/Download/?utm_source=huggingface&utm_medium=referral&utm_campaign=huggingface_repo_readme) to run the .net code on. This is a very flexible and powerful feature. Use with caution!
### Final Word
I fund the servers used to create these model files, run the Quantum Network Monitor service, and pay for inference from Novita and OpenAI—all out of my own pocket. All the code behind the model creation and the Quantum Network Monitor project is [open source](https://github.com/Mungert69). Feel free to use whatever you find helpful.
If you appreciate the work, please consider [buying me a coffee](https://www.buymeacoffee.com/mahadeva) ☕. Your support helps cover service costs and allows me to raise token limits for everyone.
I'm also open to job opportunities or sponsorship.
Thank you! 😊
- Downloads last month
- 320
Model tree for Mungert/Holo2-30B-A3B-GGUF
Base model
Qwen/Qwen3-VL-30B-A3B-Thinking
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="Mungert/Holo2-30B-A3B-GGUF", filename="", )