🦉 A suite of Whisper-style models from 250M to 18B parameters. Trained on up to 360K hours of data.