Qwen3-VL-8B-Instruct-W4A16-AutoRound-AWQ

Model Overview

This model is the AWQ (Activation-aware Weight Quantization) export of Qwen/Qwen3-VL-8B-Instruct.

It combines the speed of AWQ with the accuracy of AutoRound. The weights were fine-tuned for 800 steps to ensure the 4-bit degradation is negligible. The vision tower remains in full precision (FP16) to maintain top-tier performance on OCR and visual tasks.

Key Features

• High Performance: Optimized for vLLM serving and Transformers.
• Low VRAM: Fits comfortably on 12GB+ GPUs (RTX 3060/4070).
• Accurate Vision: Vision encoder is NOT quantized, preserving full visual acuity.

Usage

Option A: vLLM (Recommended for Speed)

pip install vllm

from vllm import LLM, SamplingParams

model_id = "Vishva007/Qwen3-VL-8B-Instruct-W4A16-AutoRound-AWQ"

llm = LLM(
    model=model_id,
    quantization="awq",
    trust_remote_code=True,
    max_model_len=4096
)

# ... (Standard vLLM inference code)

Option B: Transformers (Standard)

pip install autoawq transformers

from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
import torch

model_id = "Vishva007/Qwen3-VL-8B-Instruct-W4A16-AutoRound-AWQ"

# Load with Flash Attention 2 for best performance
model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    attn_implementation="flash_attention_2"
)
processor = AutoProcessor.from_pretrained(model_id)

# Inference Example
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"},
            {"type": "text", "text": "What does this image show?"},
        ],
    }
]

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=128)
print(processor.batch_decode(generated_ids, skip_special_tokens=True))

Citation

@misc{qwen3technicalreport,
      title={Qwen3 Technical Report}, 
      author={Qwen Team},
      year={2025}
}