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

zeromodels/qwen3-14b-base

Pure-Keras 3 conversion of Qwen/Qwen3-14B-Base for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX. This is a dense Qwen3; weights are stored in bfloat16.

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

Paper: Qwen3 Technical Report (arXiv:2505.09388) · HF Papers

✨ Quick start

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

from zeromodels.models.qwen3 import Qwen3TextGenerate, Qwen3Tokenizer

model = Qwen3TextGenerate.from_weights("zeromodels/qwen3-14b-base")
tokenizer = Qwen3Tokenizer.from_weights("zeromodels/qwen3-14b-base")

inputs = tokenizer("Give me a short introduction to large language models.")
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0]))

Load any Qwen3 variant the same way with from_weights("zeromodels/<variant>"):

Variant	Hub
`qwen3-0.6b`	zeromodels/qwen3-0.6b
`qwen3-0.6b-base`	zeromodels/qwen3-0.6b-base
`qwen3-1.7b`	zeromodels/qwen3-1.7b
`qwen3-1.7b-base`	zeromodels/qwen3-1.7b-base
`qwen3-4b`	zeromodels/qwen3-4b
`qwen3-4b-base`	zeromodels/qwen3-4b-base
`qwen3-8b`	zeromodels/qwen3-8b
`qwen3-8b-base`	zeromodels/qwen3-8b-base
`qwen3-14b`	zeromodels/qwen3-14b
`qwen3-14b-base`	zeromodels/qwen3-14b-base
`qwen3-32b`	zeromodels/qwen3-32b
`qwen3-30b-a3b`	zeromodels/qwen3-30b-a3b
`qwen3-30b-a3b-instruct-2507`	zeromodels/qwen3-30b-a3b-instruct-2507