Holo1 3B GGUF

🚀 Holo1-3B GGUF Models

This model is designed for visual document retrieval, offering a multimodal solution that combines different types of data, such as text and images. It can be used in various actions and by agents, providing a cost - efficient option for web - related tasks.

🚀 Quick Start

Holo1 models are based on the Qwen2.5 - VL architecture, which comes with transformers support. You can load the model and the processor as follows:

import json
import os
from typing import Any, Literal

from transformers import AutoModelForImageTextToText, AutoProcessor

# default: Load the model on the available device(s)
# We recommend enabling flash_attention_2 for better acceleration and memory saving.
model = AutoModelForImageTextToText.from_pretrained(
    "Hcompany/Holo1-3B",
    torch_dtype="auto",
    # torch_dtype=torch.bfloat16,
    # attn_implementation="flash_attention_2",
    device_map="auto",
)

# default processor
processor = AutoProcessor.from_pretrained("Hcompany/Holo1-3B")
# The default range for the number of visual tokens per image in the model is 4-1280.
# You can set min_pixels and max_pixels according to your needs, such as a token range of 256-1280, to balance performance and cost.
# processor = AutoProcessor.from_pretrained(model_dir, min_pixels=min_pixels, max_pixels=max_pixels)

# Helper function to run inference
def run_inference(messages: list[dict[str, Any]]) -> str:
    # Preparation for inference
    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = processor(
        text=[text],
        images=image,
        padding=True,
        return_tensors="pt",
    )
    inputs = inputs.to("cuda")

    generated_ids = model.generate(**inputs, max_new_tokens=128)
    generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
    return processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)

✨ Features

• Multimodal Capability: Supports multimodal data, including text and images, suitable for visual document retrieval.
• Cost - Efficient: Offers a cost - efficient solution for web agents, as shown in the Surfer - H benchmark.
• High - Performance Quantization: Utilizes ultra - low - bit quantization methods to balance accuracy and memory usage.
• Flexible Model Selection: Provides multiple model formats to meet different hardware and memory requirements.

📦 Installation

Since the model is based on the transformers library, you can install the necessary dependencies using pip:

pip install transformers

💻 Usage Examples

Basic Usage

import json
import os
from typing import Any, Literal

from transformers import AutoModelForImageTextToText, AutoProcessor

# default: Load the model on the available device(s)
# We recommend enabling flash_attention_2 for better acceleration and memory saving.
model = AutoModelForImageTextToText.from_pretrained(
    "Hcompany/Holo1-3B",
    torch_dtype="auto",
    # torch_dtype=torch.bfloat16,
    # attn_implementation="flash_attention_2",
    device_map="auto",
)

# default processor
processor = AutoProcessor.from_pretrained("Hcompany/Holo1-3B")
# The default range for the number of visual tokens per image in the model is 4-1280.
# You can set min_pixels and max_pixels according to your needs, such as a token range of 256-1280, to balance performance and cost.
# processor = AutoProcessor.from_pretrained(model_dir, min_pixels=min_pixels, max_pixels=max_pixels)

# Helper function to run inference
def run_inference(messages: list[dict[str, Any]]) -> str:
    # Preparation for inference
    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = processor(
        text=[text],
        images=image,
        padding=True,
        return_tensors="pt",
    )
    inputs = inputs.to("cuda")

    generated_ids = model.generate(**inputs, max_new_tokens=128)
    generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
    return processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)

Advanced Usage

Navigation with Structured Output

import json
from . import navigation

task = "Book a hotel in Paris on August 3rd for 3 nights"
prompt = navigation.get_navigation_prompt(task, image, step=1)
navigation_str = run_inference(prompt)[0]
navigation = navigation.NavigationStep(**json.loads(navigation_str))
print(navigation)
# Expected NavigationStep(note='', thought='I need to select the check-out date as August 3rd and then proceed to search for hotels.', action=ClickElementAction(action='click_element', element='August 3rd on the calendar', x=777, y=282))

Localization with click(x, y)

from . import localization

instruction = "Select July 14th as the check-out date"
prompt = localization.get_localization_prompt(image, instruction)
coordinates = run_inference(prompt)[0]
print(coordinates)
# Expected Click(352, 348)

Localization with Structured Output

import json
from . import localization

instruction = "Select July 14th as the check-out date"
prompt = localization.get_localization_prompt_structured_output(image, instruction)
coordinates_structured_str = run_inference(prompt)[0]
coordinates_structured = localization.ClickAction(**json.loads(coordinates_structured_str))
print(coordinates_structured)
# Expected ClickAction(action='click', x=352, y=340)

📚 Documentation

Model Generation Details

This model was generated using llama.cpp at commit 71bdbdb5.

Ultra - Low - Bit Quantization with IQ - DynamicGate (1 - 2 bit)

Our latest quantization method introduces precision - adaptive quantization for ultra - low - bit models (1 - 2 bit), with benchmark - proven improvements on Llama - 3 - 8B. This approach uses layer - specific strategies to preserve accuracy while maintaining extreme memory efficiency.

Benchmark Context

All tests conducted on Llama - 3 - 8B - Instruct using:

• Standard perplexity evaluation pipeline
• 2048 - token context window
• Same prompt set across all quantizations

Method

• Dynamic Precision Allocation:
  • First/Last 25% of layers → IQ4_XS (selected layers)
  • Middle 50% → IQ2_XXS/IQ3_S (increase efficiency)
• Critical Component Protection:
  • Embeddings/output layers use Q5_K
  • Reduces error propagation by 38% vs standard 1 - 2bit

Quantization Performance Comparison (Llama - 3 - 8B)

Key:

• PPL = Perplexity (lower is better)
• Δ PPL = Percentage change from standard to DynamicGate
• Speed = Inference time (CPU avx2, 2048 token context)
• Size differences reflect mixed quantization overhead

Key Improvements:

• IQ1_M shows a massive 43.9% perplexity reduction (27.46 → 15.41)
• IQ2_S cuts perplexity by 36.9% while adding only 0.2GB
• IQ1_S maintains 39.7% better accuracy despite 1 - bit quantization

Tradeoffs:

• All variants have modest size increases (0.1 - 0.3GB)
• Inference speeds remain comparable (<5% difference)

When to Use These Models

• Fitting models into GPU VRAM
• Memory - constrained deployments
• CPU and Edge Devices where 1 - 2bit errors can be tolerated
• Research into ultra - low - bit quantization

Choosing the Right Model Format

Selecting the correct model format depends on your hardware capabilities and memory constraints.

BF16 (Brain Float 16) – Use if BF16 acceleration is available

• A 16 - bit floating - point format designed for faster computation while retaining good precision.
• Provides similar dynamic range as FP32 but with lower memory usage.
• Recommended if your hardware supports BF16 acceleration (check your device's specs).
• Ideal for high - performance inference with reduced memory footprint compared to FP32.

Use BF16 if:

• Your hardware has native BF16 support (e.g., newer GPUs, TPUs).
• You want higher precision while saving memory.
• You plan to requantize the model into another format.

Avoid BF16 if:

• Your hardware does not support BF16 (it may fall back to FP32 and run slower).
• You need compatibility with older devices that lack BF16 optimization.

F16 (Float 16) – More widely supported than BF16

• A 16 - bit floating - point format with high precision but a smaller range of values than BF16.
• Works on most devices with FP16 acceleration support (including many GPUs and some CPUs).
• Slightly lower numerical precision than BF16 but generally sufficient for inference.

Use F16 if:

• Your hardware supports FP16 but not BF16.
• You need a balance between speed, memory usage, and accuracy.
• You are running on a GPU or another device optimized for FP16 computations.

Avoid F16 if:

• Your device lacks native FP16 support (it may run slower than expected).
• You have memory limitations.

Quantized Models (Q4_K, Q6_K, Q8, etc.) – For CPU & Low - VRAM Inference

Quantization reduces model size and memory usage while maintaining as much accuracy as possible.

• Lower - bit models (Q4_K) → Best for minimal memory usage, may have lower precision.
• Higher - bit models (Q6_K, Q8_0) → Better accuracy, requires more memory.

Use Quantized Models if:

• You are running inference on a CPU and need an optimized model.
• Your device has low VRAM and cannot load full - precision models.
• You want to reduce memory footprint while keeping reasonable accuracy.

Avoid Quantized Models if:

• You need maximum accuracy (full - precision models are better for this).
• Your hardware has enough VRAM for higher - precision formats (BF16/F16).

Very Low - Bit Quantization (IQ3_XS, IQ3_S, IQ3_M, Q4_K, Q4_0)

These models are optimized for extreme memory efficiency, making them ideal for low - power devices or large - scale deployments where memory is a critical constraint.

• IQ3_XS: Ultra - low - bit quantization (3 - bit) with extreme memory efficiency.

  • Use case: Best for ultra - low - memory devices where even Q4_K is too large.
  • Trade - off: Lower accuracy compared to higher - bit quantizations.

• IQ3_S: Small block size for maximum memory efficiency.

  • Use case: Best for low - memory devices where IQ3_XS is too aggressive.

• IQ3_M: Medium block size for better accuracy than IQ3_S.

  • Use case: Suitable for low - memory devices where IQ3_S is too limiting.

• Q4_K: 4 - bit quantization with block - wise optimization for better accuracy.

  • Use case: Best for low - memory devices where Q6_K is too large.

• Q4_0: Pure 4 - bit quantization, optimized for ARM devices.

  • Use case: Best for ARM - based devices or low - memory environments.

Summary Table: Model Format Selection

Model Description

Holo1 is an Action Vision - Language Model (VLM) developed by HCompany for use in the Surfer - H web agent system. It is designed to interact with web interfaces like a human user.

As part of a broader agentic architecture, Holo1 acts as a policy, localizer, or validator, helping the agent understand and act in digital environments.

Trained on a mix of open - access, synthetic, and self - generated data, Holo1 enables state - of - the - art (SOTA) performance on the WebVoyager benchmark, offering the best accuracy/cost tradeoff among current models. It also excels in UI localization tasks such as Screenspot, Screenspot - V2, [Screenspot - Pro](https://huggingface.co/datasets/likaixin/ScreenSpot - Pro), [GroundUI - Web](https://huggingface.co/datasets/agent - studio/GroundUI - 1K), and the newly introduced benchmark WebClick.

Holo1 is optimized for both accuracy and cost - efficiency, making it a strong open - source alternative to existing VLMs.

For more details, check the paper and the blog post.

• Developed by: HCompany
• Model type: Action Vision - Language Model
• Finetuned from model: Qwen/Qwen2.5 - VL - 3B - Instruct
• Paper: https://arxiv.org/abs/2506.02865
• Blog Post: https://www.hcompany.ai/surfer - h
• License: https://huggingface.co/Hcompany/Holo1 - 3B/blob/main/LICENSE

Results

Surfer - H: Pareto - Optimal Performance on WebVoyager

Surfer - H is designed to be flexible and modular. It is composed of three independent components:

• A Policy model that plans, decides, and drives the agent's behavior
• A Localizer model that sees and understands visual UIs to drive precise interactions
• A Validator model that checks whether the answer is valid

The agent thinks before acting, takes notes, and can retry if its answer is rejected. It can operate with different models for each module, allowing for tradeoffs between accuracy, speed, and cost.

The team evaluated Surfer - H on the WebVoyager benchmark: 643 real - world web tasks ranging from retrieving prices to finding news or scheduling events.

The team tested multiple configurations, from GPT - 4 - powered agents to 100% open Holo1 setups. Among them, the fully Holo1 - based agents offered the strongest tradeoff between accuracy and cost:

• Surfer - H + Holo1 - 7B: 92.2% accuracy at $0.13 per task
• Surfer - H + GPT - 4.1: 92.0% at $0.54 per task
• Surfer - H + Holo1 - 3B: 89.7% at $0.11 per task
• Surfer - H + GPT - 4.1 - mini: 88.8% at $0.26 per task

This places Holo1 - powered agents on the Pareto frontier, delivering the best accuracy per dollar. Unlike other agents that rely on custom APIs or brittle wrappers, Surfer - H operates purely through the browser – just like a real user. Combined with Holo1, it becomes a powerful, general - purpose, cost - efficient web automation system.

Holo1: State - of - the - Art UI Localization

A key skill for the real - world utility of VLMs within agents is localization: the ability to identify precise coordinates on a user interface (UI) to interact with to complete a task or follow an instruction. To assess this capability, the team evaluated the Holo1 models on several established localization benchmarks, including Screenspot, Screenspot - V2, [Screenspot - Pro](https://huggingface.co/datasets/likaixin/ScreenSpot - Pro), [GroundUI - Web](https://huggingface.co/datasets/agent - studio/GroundUI - 1K), and the newly introduced benchmark WebClick.

🔧 Technical Details

The model is based on the Qwen2.5 - VL architecture and is fine - tuned from the Qwen/Qwen2.5 - VL - 3B - Instruct model. The quantization methods used in the model, such as IQ - DynamicGate, are designed to balance accuracy and memory usage. The model is trained on a mix of open - access, synthetic, and self - generated data to achieve state - of - the - art performance on relevant benchmarks.

📄 License

The model is licensed under the terms specified in [https://huggingface.co/Hcompany/Holo1 - 3B/blob/main/LICENSE](https://huggingface.co/Hcompany/Holo1 - 3B/blob/main/LICENSE).

⚠️ Important Note

Holo1 is using absolute coordinates (number of pixels) and HuggingFace processor is doing image resize. To have matching coordinates, one needs to smart_resize the image.

💡 Usage Tip

We recommend enabling flash_attention_2 when loading the model for better acceleration and memory saving.