Run Qwen3.5 with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/qwen3.5-2b-base

Pure-Keras 3 conversion of Qwen/Qwen3.5-2B-Base for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX. Qwen3.5 is a native vision-language model: a Qwen3-VL vision tower + a dense Gated-DeltaNet / gated-full-attention hybrid text decoder. Weights are stored in bfloat16.

This repo hosts the full VLM. Load it with Qwen3_5ConditionalGenerate for image + text, or with Qwen3_5TextGenerate for text-only (it reads just the language model and drops the vision tower, like transformers' Qwen3_5ForCausalLM).

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

✨ Quick start

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

# --- text-only (loads just the language model, like transformers' Qwen3_5ForCausalLM) ---
from zeromodels.models.qwen3_5 import Qwen3_5TextGenerate, Qwen3_5Tokenizer

model = Qwen3_5TextGenerate.from_weights("zeromodels/qwen3.5-2b-base")
tokenizer = Qwen3_5Tokenizer.from_weights("zeromodels/qwen3.5-2b-base")
outputs = model.generate(**tokenizer("The capital of France is"), max_new_tokens=32)
print(tokenizer.decode(outputs[0]))

# --- image + text (the full vision-language model) ---
from PIL import Image
from zeromodels.models.qwen3_5 import Qwen3_5ConditionalGenerate, Qwen3_5Processor

model = Qwen3_5ConditionalGenerate.from_weights("zeromodels/qwen3.5-2b-base")
processor = Qwen3_5Processor.from_weights("zeromodels/qwen3.5-2b-base")
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]))

Special Thanks

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

License: Apache 2.0.