Kimi-Audio

Moonshot AI positions Kimi-Audio as a single foundation model for essentially every audio task: speech recognition, audio question answering, captioning, emotion recognition, sound and scene classification, and full end-to-end speech conversation. The architecture has three parts, a hybrid tokenizer producing discrete semantic tokens alongside continuous acoustic features, a 7B parameter audio LLM initialized from Qwen2.5-7B whose parallel heads generate text and audio tokens at once, and a flow-matching detokenizer paired with a vocoder for streaming waveform output. Pretraining covered more than 13 million hours of speech, music, and general sound, and at its April 2025 release it reported leading scores across audio understanding benchmarks. The repository provides inference code, a Docker image, and fine-tuning examples, with both instruct and pretrained checkpoints on Hugging Face, and running the 7B model locally requires a CUDA GPU. Licensing mixes Apache-2.0 for the Qwen-derived portions with MIT for the remainder, keeping commercial use open, which makes it a practical base for teams building speech-native assistants.