Whisper 3 Large is an automatic speech recognition and speech translation model built on a Transformer encoder-decoder architecture with 1.55 billion parameters, trained on more than five million hours of audio. It is multilingual-only, with no separate English-only variant, processes audio in 30-second windows, and supports up to 99 languages, making it a strong general-purpose option for teams that need a single ASR system across diverse locales and dialects.

The model weights are freely available under the MIT license, so the same weights can run through a hosted API or be deployed fully on-device for self-hosted workflows with privacy or latency requirements. In head-to-head capability comparisons, Whisper 3 Large is characterized as an audio-in, text-out specialist without general reasoning or coding abilities, and recent industry coverage notes that newer entrants such as Microsoft's MAI-Transcribe-1 have surpassed it on the FLEURS benchmark across eleven core languages, which is worth weighing for projects that prioritize top-tier transcription accuracy over open-weight flexibility.