Converted on 2026-06-11 with dsainvg/openvino-model-conv via qwen35/scripts/convert_to_openvino.py — pure ov.convert_model on the PyTorch model (no optimum-intel exporter).

Load with OpenVINO

python

import openvino as ov
import numpy as np
from transformers import AutoTokenizer

core     = ov.Core()
compiled = core.compile_model("openvino_model.xml", "CPU")
tok      = AutoTokenizer.from_pretrained("durgasai299792458/qwen35-4b-openvino-split-ir")

inputs    = tok("Hello, world!", return_tensors="np")
# Build state tensors (zeros) matching your max_seq budget, then call compiled(...)

Files

File	Description
`openvino_model.xml` / `.bin`	OpenVINO IR (FP16 weights)
`tokenizer.json`, `tokenizer_config.json`	Tokeniser assets
`config.json`, `generation_config.json`	Model config

durgasai299792458/QWEN-3.5-4B-OPENVINO-NPU-MODEL

Qwen3.5-4B — OpenVINO IR (FP16)

Load with OpenVINO

Files

durgasai299792458/
QWEN-3.5-4B-OPENVINO-NPU-MODEL