Ming-3.0-audio-flash

Ming-3.0-audio-flash is a native audio-language model from inclusionAI. It accepts audio together with text instructions and supports speech recognition, spoken question answering, audio understanding, audio reasoning, and optional synthesized-speech responses.

Model Summary

Highlights

• Complex speech recognition: Handles Chinese and English speech, dialects, accents, noisy environments, and multi-speaker recordings.
• Long-audio understanding: Comprehends and analyzes long-form recordings.
• Multidimensional perception: Understands speech, non-speech sounds, acoustic scenes, and music.
• Audio reasoning: Answers questions and performs reasoning over audio content.
• Audio interaction: Produces text responses and can optionally synthesize speech with built-in voices or a reference voice.

Architecture

Ming-3.0-audio-flash combines an audio encoder and projector with a Mixture-of-Experts language backbone. An optional Talker generates speech responses from the model's output.

Quick Start with Transformers

The model uses custom Transformers code. Review the repository code before enabling trust_remote_code=True.

git clone https://github.com/inclusionAI/Ming-3.0-audio-flash.git
cd Ming-3.0-audio-flash

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

Audio Understanding

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

Speech Recognition

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

The validated ASR prompt asks the model to identify the spoken language and place a tab between the language label and transcription.

Direct Python Usage

import torch
from transformers import AutoModelForCausalLM, AutoProcessor

model_id = "inclusionAI/Ming-3.0-audio-flash"
audio_path = "examples/audio/introduce_ant_group.wav"

processor = AutoProcessor.from_pretrained(
    model_id,
    trust_remote_code=True,
    use_fast=False,
)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
).eval()

messages = [
    {
        "role": "user",
        "content": [
            {"type": "audio", "audio": audio_path},
            {"type": "text", "text": "Please answer the question in the audio."},
        ],
    }
]

inputs = processor(messages).to(model.device)
inputs["audio_feats"] = inputs["audio_feats"].to(model.dtype)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=False,
    )

generated_ids = output_ids[0, inputs["input_ids"].shape[1] :]
response = processor.decode(generated_ids, skip_special_tokens=True).strip()
print(response)

Speech Output

The Talker example supports built-in voices and reference-audio voice prompting:

python examples/transformers_thinker_talker.py \
  --model inclusionAI/Ming-3.0-audio-flash \
  --audio examples/audio/introduce_ant_group.wav \
  --voice US_Stella \
  --output output.wav

See the voice guide for available voice names and reference-audio usage.

Serving

vLLM-Omni

export MODEL_PATH=inclusionAI/Ming-3.0-audio-flash
MING_OUTPUT_MODE=text ./scripts/serve_vllm_omni.sh "$MODEL_PATH"

The reference 128K serving profile uses tensor parallelism across four 80 GB H800 GPUs. Select the deployment YAML that matches text-only or Talker output. On particularly long audio inputs, the four-GPU profiles may OOM; in that case, switch to the corresponding eight-GPU fallback profile (deploy/*_8gpu.yaml), which sets tensor_parallel_size: 8 across GPUs 0–7 (and the Talker reuses GPU 7). See the deployment guide for details.

For speech output, start the audio profile and select an IP or registered voice with the vLLM-Omni client's --voice option:

MING_OUTPUT_MODE=audio ./scripts/serve_vllm_omni.sh "$MODEL_PATH"
python examples/vllm_omni_audio_chat.py examples/audio/introduce_ant_group.wav \
  --audio-response \
  --voice US_Stella \
  --output-audio outputs/answer.wav

The client sends this selection as Chat Completions audio.voice: {"id": "..."}. vLLM-Omni uses the built-in IP voice 灵小柔 when no voice is specified. The model directory also provides 灵楷 and US_James as preset reference voices under voice_presets/voice_audio/, with their matching transcripts in voice_presets/voice_prompts.json. Both names are also available as built-in IP voices.

SGLang

export MODEL_PATH=inclusionAI/Ming-3.0-audio-flash
./scripts/serve_sglang.sh "$MODEL_PATH"

The validated SGLang profile uses four GPUs and provides text output. See the SGLang example for client usage.

SGLang-Omni

SGLang-Omni serves transcription, audio chat, and speech synthesis from one server. Serving uses two GPUs: --thinker.tp_size 2 places the thinker on both visible devices, the audio encoder uses GPU 0, and speech synthesis uses GPU 1.