Ming-3.0-audio-tiny

Ming-3.0-audio-tiny

Listens to audio with a text instruction and answers in text, with optional spoken replies. A compact audio-language model from inclusionAI for speech recognition, spoken questions and audio understanding in Chinese and English.

Made by inclusionAI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Ming-3.0-audio-tiny comes from inclusionAI and handles audio together with text. You give it audio plus a text instruction, and it does speech recognition, answers spoken questions, understands sounds and reasons about audio content. It replies in text, and an optional Talker part can synthesize a spoken reply. The card puts it at 9.3B total parameters with 2.7B active per step.

Who it suits: Developers and researchers who work with Chinese or English audio, including dialects, accents and noisy recordings, and who can run custom Transformers code on a GPU. The files carry the MIT license.

Industry | Who uses it | How they use it | Example
Media | Transcription teams | Transcribe speech in noisy or multi-speaker recordings | Turning a panel recording into text
Customer service | Contact-center developers | Ask questions about a recorded call | Asking what a caller wanted
Research | Audio researchers | Test understanding of speech, sounds, scenes and music | Asking what kind of place a clip was recorded in

Anyone who only needs plain text chat can skip it, since this model is built around audio input.

Where to use it:

• Python with the transformers library and the model's custom code
• A Linux or similar shell setup, as the card's steps use a virtual environment
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Speech recognition | Developers | A transcript, with the spoken language named and a tab before the text
Audio question answering | App builders | A text answer to a question asked in the audio
Sound and music understanding | Researchers | Text describing the sounds, scene or music

How to set it up:

• Python and a virtual environment
• The example repository, Ming-3.0-audio-flash, cloned from GitHub
• The requirements file for Transformers in that repository

1. Clone the Ming-3.0-audio-flash repository from GitHub and open its folder.
2. Create a virtual environment, activate it and install the requirements with the commands below.
3. Review the repository's code before you enable trust_remote_code=True, as the card advises.

cd Ming-3.0-audio-flash

python -m venv .venv-transformers
source .venv-transformers/bin/activate
pip install -r requirements-transformers.txt

How to use it: You run the example script with an audio file, the model name and a prompt. For speech recognition, add the asr task, a token cap and a repeat limit.

python examples/transformers_audio.py examples/audio/introduce_ant_group.wav \
  --model inclusionAI/Ming-3.0-audio-tiny \
  --prompt "Please answer the question in the audio."

python examples/transformers_audio.py examples/audio/introduce_ant_group.wav \
  --model inclusionAI/Ming-3.0-audio-tiny \
  --task asr \
  --max-new-tokens 256 \
  --no-repeat-ngram-size 10

Tool | What it does
examples/transformers_audio.py | Runs the model on an audio file
--prompt | Gives the text instruction that goes with the audio
--task asr | Switches to speech recognition
--max-new-tokens | Caps how many tokens the reply can have

The validated recognition prompt asks the model to name the spoken language and put a tab between that label and the transcript.

Example: You run the first command on the bundled recording. The model hears the spoken question and writes its answer as text.

For newcomers and experts: A good first step is the first command above on the bundled example audio, since it needs no extra settings. Read the answer, then try the recognition command.

The model joins an audio encoder and projector to a compact mixture-of-experts language backbone. The optional Talker turns the model's output into speech, using built-in voices or a reference voice.

Set up the environment
Pass audio and a text instruction
The encoder and language model respond
Get a text answer
Optionally synthesize speech

Good to know:

• License: MIT
• Cost to run: none named
• Data it reaches: the audio files you give it, and the download from Hugging Face
• If you use a reference voice for spoken replies, use only voices you have the right to use
• Hugging Face lists 8.9B parameters, while the card's table says 9.3B total

On Hugging Face, 13 November 2026

• Task: audio to audio
• Library: transformers
• Parameters: 8.9B
• Weights: about 17.7 GB in BF16; about 4.4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 5,202
• Likes: 62
• License: mit
• Languages: en, zh
• Quantized versions: 4
• Fine-tunes: 2
• Tags: transformers, onnx, safetensors, bailing_moe_v3_audio, feature-extraction, audio, speech, automatic-speech-recognition, audio-understanding, audio-reasoning, speech-to-text, text-to-speech

Links

• Model: https://huggingface.co/inclusionAI/Ming-3.0-audio-tiny
• Files: https://huggingface.co/inclusionAI/Ming-3.0-audio-tiny/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download inclusionAI/Ming-3.0-audio-tiny

Then follow the usage in its model card. Nothing in this kit runs by itself.