Property	Value
Base model	Qwen/Qwen3.5-0.8B
Architecture	Qwen3.5 (hybrid linear-attention + full-attention)
Weight format	NF4 (4-bit NormalFloat)
Quantization	Symmetric, channel-wise (group_size=-1)
Shapes	Dynamic
State management	Stateful (KV cache + conv/recurrent states)

Architecture

Qwen3.5 is a hybrid model with 24 layers: 18 linear-attention (GatedDeltaNet) + 6 full-attention layers in a 3:1 repeating pattern. Linear-attention layers use fixed-size conv and recurrent states (cache_params.past.{conv,ssm}.N).

Conversion

bash

pip install git+https://github.com/rkazants/optimum-intel.git@support_qwen3_5
pip install --pre -U openvino openvino-tokenizers nncf --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly
pip install transformers==5.2.0

optimum-cli export openvino \
  --model Qwen/Qwen3.5-0.8B \
  --weight-format nf4 \
  --group-size -1 \
  Qwen3.5-0.8B_nf4_sym_group-1_dyn_stateful

dylanneve1/Qwen3.5-0.8B_nf4_sym_group-1_dyn_stateful

Qwen3.5-0.8B — NF4 Channel-wise (OpenVINO)

Model Details

Architecture

Conversion

dylanneve1/
Qwen3.5-0.8B_nf4_sym_group-1_dyn_stateful