Kimi-Audio-7B-Instruct
Kimi Audio Instruct supports recognition, audio questions and spoken conversation. It uses the Qwen2.5-7B-derived audio architecture after task-oriented fine-tuning, rather than the separate pretrained-only release.
One model supports several audio tasks
The card demonstrates transcription from an audio file and generation of a spoken answer with its text. Task support also includes audio captioning, emotion recognition and sound-event classification. English and Chinese appear in the published model metadata.Kimi-Audio-7B-Instruct model card
Waveform generation needs the audio components
The implementation converts generated semantic tokens into waveforms through a flow-matching detokenizer and BigVGAN vocoder. Its example can load the detokenizer only when audio output is needed, and writes the generated waveform at 24 kHz.Kimi Audio implementation and dated release history
Separate transcript quality from speaking speed
Practical guidance. For a voice interface, measure recognition errors, answer quality and time to audible output separately. A correct transcript does not establish that a spoken answer is correct or that the complete conversation meets your latency target.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | moonshotai/Kimi-Audio-7B-Instruct |
| Training ancestry | Qwen2.5-7B-based Kimi Audio, followed by instruction fine-tuning.Kimi-Audio-7B-Instruct model card |
| Checkpoint license | MIT in the model metadata; Qwen-derived code retains Apache 2.0 terms.Kimi-Audio-7B-Instruct model card |
| Output modes | Text only, or audio with text.Kimi-Audio-7B-Instruct model card |
Related models
- Kimi-Audio-7B
Moonshot AI ยท Kimi Audio
Pretrained audio-language model for adaptation to speech, sound and audio conversation tasks.
Launch date Downloadable