Qwen2.5-VL-7B-Instruct-ternary

Qwen2.5-VL-7B-Instruct — Ternary Quantized

Ternary-quantized version of Qwen/Qwen2.5-VL-7B-Instruct, produced with ternary-quant.

Qwen2.5-VL is one of the most capable open-weight vision-language models. This ternary-quantized version targets lower memory use while preserving multimodal behavior in a HuggingFace-native, component-aware PTQ workflow.

Quantization details

Usage

from ternary_quant.inference import load_ternary_model

model, processor = load_ternary_model(
    "AsadIsmail/Qwen2.5-VL-7B-Instruct-ternary",
    runtime_mode="metal"  # "cached" for NVIDIA/CPU
)

from PIL import Image
image = Image.open("photo.jpg")
inputs = processor(text="What is shown in this image?", images=image, return_tensors="pt")
inputs = {k: v.to(model.device) for k, v in inputs.items()}
outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0], skip_special_tokens=True))

Reproduce

pip install ternary-quant
ternary-quant quantize-broad Qwen/Qwen2.5-VL-7B-Instruct \
    --output ./Qwen2.5-VL-7B-Instruct-ternary \
    --components text_backbone multimodal_connector \
    --scheme tritplane3 --dtype float16 --eval

Part of the ternary-models collection

github.com/Asad-Ismail/ternary-models