Kimi-Audio-7B
Kimi Audio 7B is the pretrained checkpoint in Moonshot's audio model family. It extends Qwen2.5-7B through audio and text training. The publisher directs users who need immediate task responses to Instruct.
Training extends beyond speech transcripts
The research uses more than 13 million hours of audio, including speech, environmental sounds and music, together with text. It combines continuous input features with discrete output tokens and applies several tasks during continued pretraining.Kimi Audio technical report
Two representations preserve different information
The implementation combines semantic audio tokens with acoustic features from a Whisper encoder. Both run at 12.5 frames per second after processing. Shared language-model layers then feed separate text and audio output heads.Kimi Audio implementation and dated release history
Select the checkpoint for the training stage
Practical guidance. This release is a starting point for adaptation. Do not treat the base checkpoint as interchangeable with Instruct when measuring recognition or conversation quality. Keep your task data, output format and evaluation target explicit.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | moonshotai/Kimi-Audio-7B |
| Training ancestry | Qwen2.5-7B, modified for audio and text.Kimi-Audio-7B model card |
| Checkpoint license | MIT in the model metadata; Qwen-derived code retains Apache 2.0 terms.Kimi-Audio-7B model card |
| Training stage | Pretrained checkpoint for further fine-tuning.Kimi-Audio-7B model card |
Related models
- Kimi-Audio-7B-Instruct
Moonshot AI ยท Kimi Audio
Instruction-tuned audio model that can answer in text or produce speech with its text response.
Launch date Downloadable