OFA framework efficiently initializes unseen subword embeddings for large-scale multilingual continued pretraining, leading to improved performance and faster convergence.