Microsoft is open-sourcing an optimized version of Google's BERT that uses ONNX Runtime and CPUs or GPUs to speed language model performance.