Try Kimi at kimi.com →

Kimi-Audio
Moonshot's Open Audio Model

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.

Open source 7B parameters 13M+ hours of audio
AUDIO IN speech, music, sound TOKENIZER12.5 Hz discrete semantic tokens continuous Whisper features K AudioLLM 7Bfrom Qwen2.5-7B TEXT OUT “Sure, here’s theweather for today…” transcripts, answers SPEECH OUTstreaming detokenizer flow matching + vocoder24 kHz audio, in chunks One model for listening, understanding and talking. Simplified diagram.
OVERVIEW

Kimi-Audio at a glance

7BParameters

A compact language model at the core, built on Qwen2.5-7B.

13M+Hours of audio

Speech, music and everyday sounds used in pre-training.

1.28%LibriSpeech error

Word error rate on test-clean, Moonshot-reported.

MITOpen licence

Code, weights and an evaluation toolkit are all public.

Good to know: Kimi-Audio is a separate research model, not a feature inside the Kimi app. You download it and run it on your own hardware.
SPECS

Kimi-Audio specifications

SpecKimi-Audio
Released25 April 2025 (Instruct model, code, report); 27 April 2025 (base model)
DeveloperMoonshot AI (Kimi Team)
ModelsKimi-Audio-7B (base) and Kimi-Audio-7B-Instruct
Language modelAbout 7B parameters, initialised from Qwen2.5-7B
Training dataMore than 13 million hours of speech, music and sound, plus text
Audio inputDiscrete semantic tokens plus continuous Whisper features, both at 12.5 Hz
Audio outputStreaming flow-matching detokenizer and vocoder, 24 kHz speech
TasksSpeech recognition, audio Q&A, captioning, emotion and sound recognition, voice chat
LicenceMIT for Moonshot's code; Apache 2.0 for code derived from Qwen2.5-7B
WhereGitHub (MoonshotAI/Kimi-Audio) and Hugging Face (moonshotai)
Hosted APINone. Run it yourself.

From the Kimi-Audio GitHub repository and technical report.

TASKS

What Kimi-Audio can do

Speech to text

Transcribe English and Mandarin recordings with very low error rates, including meetings and noisy audio.

Answer questions about audio

Ask what's happening in a clip, what someone said or what instrument is playing.

Recognise sounds and scenes

Tell a dog bark from a doorbell, or a café from a street, from the audio alone.

Detect emotion

Pick up whether a speaker sounds happy, angry or sad.

Talk back

Hold a spoken conversation and reply in speech, with control over speed, accent and emotion.

Fine-tune it

Moonshot published a fine-tuning example for the base model, so you can adapt it to your own audio.

HOW

How Kimi-Audio works

1. Two ways of hearing

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.

2. One language model

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.

3. Speaking in chunks

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.

Training: Moonshot pre-trained on audio-only, text-only, audio-to-text and mixed audio-text tasks, then fine-tuned for instructions. It says most fine-tuning data came from openly available sources.
BENCHMARKS

Kimi-Audio benchmarks

Speech recognition (word error rate, lower is better)

TestKimi-AudioQwen2-AudioQwen2.5-Omni
LibriSpeech (clean / other)1.28 / 2.421.74 / 4.042.37 / 4.21
Fleurs (zh / en)2.69 / 4.44—2.92 / 4.17
AISHELL-10.601.521.13
AISHELL-2 iOS2.563.082.56
Kimi-ASR internal (1 / 2)1.42 / 2.442.31 / 3.241.53 / 2.68

Audio understanding and chat (higher is better)

TestKimi-AudioQwen2.5-Omni
VocalSoundVocal sounds94.8593.73
Nonspeech7kNon-speech sounds93.9369.89
MELDEmotion in speech59.1349.83
TUT2017Acoustic scenes65.2543.27
VoiceBench averageSpoken chat and instructions76.9372.83

All results from the Kimi-Audio technical report (April 2025), reproducible with Moonshot's open Kimi-Audio-Evalkit.

SPEECH RECOGNITION ERRORSWord error rate, % · lower is better LibriSpeech test-clean Kimi-Audio 1.28 Qwen2-Audio 1.74 Qwen2.5-Omni 2.37 Baichuan 3.02 AISHELL-1 (Mandarin) Kimi-Audio 0.60 Qwen2.5-Omni 1.13 Qwen2-Audio 1.52 Baichuan 1.93 From the Kimi-Audio technical report, April 2025. Rivals are 2025 models; newer audio models may do better.

Kimi-Audio had the lowest error rates in Moonshot's English and Mandarin tests.

RUN

How to run Kimi-Audio

You'll need Python and an NVIDIA GPU. Install the package, then load Kimi-Audio-7B-Instruct from Hugging Face.

# Option 1: install the package pip install torch pip install git+https://github.com/MoonshotAI/Kimi-Audio.git # Option 2: clone the repo (includes examples and fine-tuning code) git clone https://github.com/MoonshotAI/Kimi-Audio.git cd Kimi-Audio git submodule update --init --recursive pip install -r requirements.txt

Models

moonshotai/Kimi-Audio-7B-Instruct for ready-to-use tasks, moonshotai/Kimi-Audio-7B as a base for fine-tuning.

Evaluation

Kimi-Audio-Evalkit reproduces the published scores and lets you compare other audio models the same way.

Output modes

Use output_type="text" for transcripts and answers, or "both" to get speech back too.

WHICH

Kimi-Audio vs the Kimi app and K3.1 Pro

You want to…UseWhy
Transcribe or analyse audio on your own serversKimi-AudioOpen model you control, strong at speech recognition
Talk to Kimi on your phoneThe Kimi appBuilt-in voice features, no setup
Reason over long documents, code or videoKimi K3.1 ProMoonshot'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.

LIMITS

Kimi-Audio limitations

It's a 2025 model

Benchmarks compare it with early-2025 rivals. Newer audio models may beat it, so test on your own audio.

Chinese-first speech output

Moonshot's voice-generation test set is mostly Mandarin. English speech output hasn't been evaluated as thoroughly.

You need a GPU

There's no hosted API. You'll need a capable NVIDIA GPU and some Python setup to run it.

No newer version yet

As of November 2026 Moonshot hasn't released a Kimi-Audio successor, and its K3 models don't take audio files.

FAQ

Kimi-Audio FAQ

Kimi-Audio is an open-source audio foundation model from Moonshot AI, the company behind Kimi. Released in April 2025, it can transcribe speech, answer questions about audio, recognise sounds and emotions, and hold spoken conversations, all in one 7B-parameter model.

Yes. The code and model weights are free to download from GitHub and Hugging Face. Moonshot's code is under the MIT licence, and parts derived from Qwen2.5-7B are under Apache 2.0, which generally allows commercial use. You pay only for the hardware you run it on.

Very good for its size. Moonshot reports a 1.28% word error rate on LibriSpeech test-clean and 0.60% on the Mandarin AISHELL-1 test, ahead of Qwen2-Audio and Qwen2.5-Omni in its comparisons.

Not directly. Kimi-Audio is a separate open model you run yourself. The Kimi app has its own voice features, and the Kimi API's current models accept text, images and video rather than audio files.

A machine with a capable NVIDIA GPU and Python. The core language model has about 7 billion parameters, plus the audio tokenizer and speech decoder. Check the GitHub README for current requirements.

Its published results focus on English and Mandarin Chinese. Its speech-generation tests are mainly in Mandarin, so test English voice output on your own audio first.

Yes. Moonshot released the pre-trained Kimi-Audio-7B base model and a fine-tuning example on 29 May 2025, so you can adapt it to your own data or tasks.

Not as of November 2026. Kimi-Audio-7B and Kimi-Audio-7B-Instruct from April 2025 are still Moonshot's latest open audio models.

Explore more Kimi models

From open audio and coding models to the K3.1 Pro flagship.