Modelpublic

xpuenabler/Qwen3.6-35B-A3B-fp16-ov

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes5downloads

Model Card

Qwen3.6-35B-A3B — OpenVINO IR (FP16)

OpenVINO IR export (weight-format FP16) of Qwen/Qwen3.6-35B-A3B (qwen3_5_moe, multimodal MoE, 256 experts / top-8, 40 layers).

Export environment

—optimum-intel @ git main · openvino==2026.2.1 · transformers==5.2.0

Export command

bash

optimum-cli export openvino \
  --model Qwen/Qwen3.6-35B-A3B \
  --task image-text-to-text \
  --trust-remote-code \
  --weight-format fp16 \
  <output_dir>

Notes

—OpenVINO IR is device-agnostic; for the OpenVINO GPU device (Intel iGPU/Arc/dGPU), FP16 is the native compute precision.
—MoE experts are stored as fused 3D weight tensors and executed via batched MatMul (not per-expert split).

Usage

python

from optimum.intel import OVModelForVisualCausalLM
from transformers import AutoProcessor
model = OVModelForVisualCausalLM.from_pretrained("xpuenabler/Qwen3.6-35B-A3B-fp16-ov", device="GPU")
processor = AutoProcessor.from_pretrained("xpuenabler/Qwen3.6-35B-A3B-fp16-ov")