qwen2.5-vl-3b-instruct

See our collection for all Qwen2.5-VL sizes.

Run Qwen2.5-VL with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/qwen2.5-vl-3b-instruct

Pure-Keras 3 conversion of Qwen/Qwen2.5-VL-3B-Instruct for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX. This is the 3B variant, served here as image + text -> text via Qwen2_5VLProcessor; weights are stored in bfloat16.

For model details, license, and usage terms, see the upstream model card.

Paper: Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution (arXiv:2409.12191) · HF Papers

Paper: YaRN: Efficient Context Window Extension of Large Language Models (arXiv:2309.00071) · HF Papers

Paper: Qwen-VL: A Frontier Large Vision-Language Model with Versatile Abilities (arXiv:2308.12966) · HF Papers

✨ Quick start

Text-only

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from zeromodels.models.qwen2_5_vl import Qwen2_5VLTextGenerate, Qwen2_5VLProcessor

model = Qwen2_5VLTextGenerate.from_weights("zeromodels/qwen2.5-vl-3b-instruct")
processor = Qwen2_5VLProcessor.from_weights("zeromodels/qwen2.5-vl-3b-instruct")

inputs = processor(conversation=[
    {"role": "user", "content": [{"type": "text", "text": "Hello, who are you?"}]}
])
outputs = model.generate(**inputs, max_new_tokens=64)
print(processor.decode(outputs[0]))

Image + text

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from PIL import Image
from zeromodels.models.qwen2_5_vl import Qwen2_5VLConditionalGenerate, Qwen2_5VLProcessor

model = Qwen2_5VLConditionalGenerate.from_weights("zeromodels/qwen2.5-vl-3b-instruct")
processor = Qwen2_5VLProcessor.from_weights("zeromodels/qwen2.5-vl-3b-instruct")

inputs = processor(conversation=[
    {"role": "user", "content": [
        {"type": "image", "image": Image.open("photo.jpg")},
        {"type": "text", "text": "Describe this image in one sentence."},
    ]}
])
outputs = model.generate(**inputs, max_new_tokens=64)
print(processor.decode(outputs[0]))

Load any Qwen2.5-VL variant the same way with from_weights("zeromodels/<variant>"):

Special Thanks

A huge thank you to the Qwen team at Alibaba for creating and releasing these models.

License: Qwen Research License (non-commercial; see the upstream license).