Kimi-Audio-7B-Instruct
Moonshot's open audio foundation model (built on Qwen2.5-7B) for speech recognition, audio understanding and end-to-end speech conversation. Moonshot offers no hosted realtime voice API, so this is the only Kimi voice option.
Overview
Best for: Research on audio understanding plus speech replies with a permissive MIT licence.
At a glance
Named 7B (Qwen2.5-7B base), about 10B params total. Chunk-wise streaming detokenizer. English and Chinese.
Speech audio
Speech at 24 kHz (output_type='both' returns text + audio)
English and Chinese.
Not documented.
Chunk-wise streaming detokenizer for low-latency generation; no figure.
Self-hosted.
Self-hosted.
Not stated on the card; about 10B BF16 params suggests 24 GB+ VRAM (own estimate).
MIT (Qwen2.5-derived code Apache-2.0).
Features
- ASR
- audio QA and captioning
- emotion and sound-event recognition
- end-to-end speech conversation
Pricing
No public price list.
No licence fee; you pay for GPU time.
n/a (self-hosted)
Free tier: Open weights
Source: huggingface.co
Setup
- CUDA GPU (24 GB-class suggested, own estimate).
- pip install the Kimi-Audio package or pull the Docker image.
- Load KimiAudio with load_detokenizer=True and call generate.
Endpoint
None (library)Authentication
n/a
Quick start python
pip install git+https://github.com/MoonshotAI/Kimi-Audio.git "transformers<5"
# or: docker pull moonshotai/kimi-audio:v0.1
# python
from kimia_infer.api.kimia import KimiAudio
model = KimiAudio(model_path="moonshotai/Kimi-Audio-7B-Instruct", load_detokenizer=True)
# wav, text = model.generate(messages, output_type="both")Written from the current docs. Check the vendor's SDK version before you ship.
Warnings
No realtime serving
Turn-based library; you build streaming, VAD and interruption.
Aging release
Released April 2025 and little updated since; newer open models (MiniCPM-o 4.5, Qwen3-Omni) are generally preferred.
EN/ZH only
Only English and Chinese are tagged.
Moonshot has no voice API
Kimi's hosted API is text/multimodal chat only; do not expect a managed realtime endpoint.
Plus 3 warnings that apply to all open models APIs. See category warnings.
Limits
- Not a full-duplex realtime server
- No hosted Moonshot realtime API found
Models and products
| Name | Status |
|---|---|
| moonshotai/Kimi-Audio-7B-Instruct | GA |
Docs and sources
Docs
Sources used
- huggingface.co/moonshotai/Kimi-Audio-7B-Instruct
- huggingface.co/api/models/moonshotai/Kimi-Audio-7B-Instruct
VRAM; exact import path and generate signature in the snippet.