Worked on enhancing quantized large language model compatibility for NPU hardware within the openvinotoolkit/nncf repository. Addressed NPU compiler support by updating the ONNX opset to version 21 and replacing the MatMulNBits operation with DequantizeLinear, ensuring quantized LLM models could execute effectively on NPU devices. Focused on model optimization and LLM compression techniques, leveraging Python and ONNX to implement these changes. The update enabled new deployment options and potential performance improvements for quantized models on specialized hardware. The work demonstrated a targeted approach to hardware compatibility, emphasizing practical integration and technical depth in model deployment workflows.