Laya now runs locally on CPU through an OpenVINO backend
You can run Laya's English checkpoint on a CPU through OpenVINO instead of PyTorch, using an already exported 405 MB int8 model or exporting your own checkpoint. The fork loads the OpenVINO model with `OVAgent`, while keeping the tokenizer and routing config alongside it, and reports 98.9% top-answer agreement with the Torch version. Install it directly from GitHub, since it is not on PyPI.