README.md

---
license: apache-2.0
base_model:
- inclusionAI/ZwZ-8B
datasets:
- inclusionAI/ZwZ-RL-VQA
- inclusionAI/ZoomBench
language:
- en
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- text-generation-inference
- F8_E4M3
- fp8
- vllm
- llm-compressor
---

![1](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/cJvpKspuxHdZNnkURe5jC.png)

# **ZwZ-8B-FP8**

> **ZwZ-8B-FP8** is an FP8-compressed evolution built on top of **inclusionAI/ZwZ-8B**. This variant leverages **BF16 · FP8 (F8_E4M3)** precision formats to significantly reduce memory footprint and improve inference efficiency while preserving the fine-grained multimodal perception strengths of the original architecture.
> The result is a highly efficient 8B vision-language model optimized for real-time, single-pass visual reasoning with enhanced hardware efficiency.

> [!important]
> FP8 (8-bit floating point) weight and activation quantization using hardware acceleration on GPUs – [FP8 W8A8](https://docs.vllm.ai/en/stable/features/quantization/fp8/). Quantization W8A8 FP8-dynamic recipe – [examples](https://github.com/vllm-project/llm-compressor/tree/main/examples/quantization_w8a8_fp8).

## About the Base Model

**ZwZ-8B** from inclusionAI is an 8B-parameter fine-grained multimodal perception vision-language model built upon Qwen3-VL-8B. It is trained using innovative **Region-to-Image Distillation (R2I)** combined with reinforcement learning to achieve state-of-the-art visual understanding in a single forward pass.

Unlike traditional VLMs that require inference-time zooming, cropping, or tool calling, ZwZ internalizes region-level perception directly into full-image reasoning.

### Key Innovations of ZwZ-8B

* **Region-to-Image Distillation (R2I)**:
  Teacher models such as Qwen3-VL-235B and GLM-4.5V generate high-fidelity VQA supervision on micro-cropped image regions with precise bounding boxes. This region-grounded supervision is distilled back into full-image context, allowing the student model to internalize fine-grained perception.

* **Single-Pass Fine-Grained Understanding**:
  Eliminates multi-step inference pipelines involving zooming, cropping, or external tool calls.

* **Strong Micro-Perception Capabilities**:

  * OCR and small-text detection
  * Object counting
  * Color and material attribute recognition
  * Structural analysis
  * Symbol and icon detection in dense scenes

* **Out-of-Distribution Generalization**:
  Demonstrates strong performance on:

  * Visual reasoning benchmarks
  * GUI agent tasks
  * AIGC detection
  * Complex real-world scenes

* **Edge-Optimized Deployment**:
  Enables real-time robotics and mobile vision applications without multi-stage inference overhead.

ZwZ is part of a broader model family spanning 4B, 7B, and 8B scales.

## What FP8 Adds

The **ZwZ-8B-FP8** variant introduces:

* **BF16 · FP8 (F8_E4M3) Compression**: Transformer Engine–based quantization reduces VRAM usage while maintaining strong perception fidelity.
* **Higher Throughput**: Improved tokens per second and image processing speed.
* **Lower Memory Footprint**: Better deployment feasibility on Hopper-class and compatible GPUs.
* **Production-Friendly Efficiency**: Ideal for real-time multimodal systems requiring compact yet powerful perception models.

## Quick Start with Transformers

```python
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch

# Load the FP8-compressed ZwZ-8B model
model = Qwen3VLForConditionalGeneration.from_pretrained(
    "prithivMLmods/ZwZ-8B-FP8",
    torch_dtype="auto",
    device_map="auto"
)

processor = AutoProcessor.from_pretrained(
    "prithivMLmods/ZwZ-8B-FP8"
)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Analyze the fine-grained details in this image."},
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

image_inputs, video_inputs = process_vision_info(messages)

inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=256)

generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]

output_text = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False
)

print(output_text)
```

## Intended Use

* Real-time multimodal perception systems
* Robotics and embodied AI
* GUI agents
* OCR-heavy and structured visual environments
* Edge deployment scenarios requiring single-pass inference

## Limitations & Risks

* FP8 requires compatible GPU architectures for optimal acceleration.
* While compression maintains strong fidelity, extremely fine-grained edge cases may show minor precision differences compared to full BF16.
* Users are responsible for ethical and lawful deployment.