Kimi-Audio is a free, open-source model from the makers of Kimi that listens, understands and talks. It handles speech recognition, audio questions, sound and emotion detection and voice chat in one 7B model. Here's how it works, how well it scores and how to run it.
A compact language model at the core, built on Qwen2.5-7B.
Speech, music and everyday sounds used in pre-training.
Word error rate on test-clean, Moonshot-reported.
Code, weights and an evaluation toolkit are all public.
| Spec | Kimi-Audio |
|---|---|
| Released | 25 April 2025 (Instruct model, code, report); 27 April 2025 (base model) |
| Developer | Moonshot AI (Kimi Team) |
| Models | Kimi-Audio-7B (base) and Kimi-Audio-7B-Instruct |
| Language model | About 7B parameters, initialised from Qwen2.5-7B |
| Training data | More than 13 million hours of speech, music and sound, plus text |
| Audio input | Discrete semantic tokens plus continuous Whisper features, both at 12.5 Hz |
| Audio output | Streaming flow-matching detokenizer and vocoder, 24 kHz speech |
| Tasks | Speech recognition, audio Q&A, captioning, emotion and sound recognition, voice chat |
| Licence | MIT for Moonshot's code; Apache 2.0 for code derived from Qwen2.5-7B |
| Where | GitHub (MoonshotAI/Kimi-Audio) and Hugging Face (moonshotai) |
| Hosted API | None. Run it yourself. |
From the Kimi-Audio GitHub repository and technical report.
Transcribe English and Mandarin recordings with very low error rates, including meetings and noisy audio.
Ask what's happening in a clip, what someone said or what instrument is playing.
Tell a dog bark from a doorbell, or a café from a street, from the audio alone.
Pick up whether a speaker sounds happy, angry or sad.
Hold a spoken conversation and reply in speech, with control over speed, accent and emotion.
Moonshot published a fine-tuning example for the base model, so you can adapt it to your own audio.
Incoming audio becomes discrete “semantic” tokens that capture what's said, plus continuous Whisper features that keep tone and sound detail. Both run at 12.5 frames per second.
A 7B model, started from Qwen2.5-7B and trained further on audio and text, reads both inputs and writes text and audio tokens side by side.
A flow-matching detokenizer and vocoder turn audio tokens into 24 kHz speech chunk by chunk, so replies can start playing before they're finished.
| Test | Kimi-Audio | Qwen2-Audio | Qwen2.5-Omni |
|---|---|---|---|
| LibriSpeech (clean / other) | 1.28 / 2.42 | 1.74 / 4.04 | 2.37 / 4.21 |
| Fleurs (zh / en) | 2.69 / 4.44 | — | 2.92 / 4.17 |
| AISHELL-1 | 0.60 | 1.52 | 1.13 |
| AISHELL-2 iOS | 2.56 | 3.08 | 2.56 |
| Kimi-ASR internal (1 / 2) | 1.42 / 2.44 | 2.31 / 3.24 | 1.53 / 2.68 |
| Test | Kimi-Audio | Qwen2.5-Omni |
|---|---|---|
| VocalSoundVocal sounds | 94.85 | 93.73 |
| Nonspeech7kNon-speech sounds | 93.93 | 69.89 |
| MELDEmotion in speech | 59.13 | 49.83 |
| TUT2017Acoustic scenes | 65.25 | 43.27 |
| VoiceBench averageSpoken chat and instructions | 76.93 | 72.83 |
All results from the Kimi-Audio technical report (April 2025), reproducible with Moonshot's open Kimi-Audio-Evalkit.
Kimi-Audio had the lowest error rates in Moonshot's English and Mandarin tests.
You'll need Python and an NVIDIA GPU. Install the package, then load Kimi-Audio-7B-Instruct from Hugging Face.
moonshotai/Kimi-Audio-7B-Instruct for ready-to-use tasks, moonshotai/Kimi-Audio-7B as a base for fine-tuning.
Kimi-Audio-Evalkit reproduces the published scores and lets you compare other audio models the same way.
Use output_type="text" for transcripts and answers, or "both" to get speech back too.
| You want to… | Use | Why |
|---|---|---|
| Transcribe or analyse audio on your own servers | Kimi-Audio | Open model you control, strong at speech recognition |
| Talk to Kimi on your phone | The Kimi app | Built-in voice features, no setup |
| Reason over long documents, code or video | Kimi K3.1 Pro | Moonshot's flagship; accepts text, images and video, not audio files |
New to Kimi? See how to use Kimi, download the app or read about Kimi K3.1 Pro.
Benchmarks compare it with early-2025 rivals. Newer audio models may beat it, so test on your own audio.
Moonshot's voice-generation test set is mostly Mandarin. English speech output hasn't been evaluated as thoroughly.
There's no hosted API. You'll need a capable NVIDIA GPU and some Python setup to run it.
As of November 2026 Moonshot hasn't released a Kimi-Audio successor, and its K3 models don't take audio files.