mlx-community/Ming-3.0-audio-tiny-8bit

Ming-3.0 Audio Tiny for MLX

Converted from inclusionAI/Ming-3.0-audio-tiny.

English and Chinese audio understanding, transcription, text chat and spoken dialogue. task='spoken_dialogue' loads the bundled Talker lazily and produces 44.1 kHz audio. Only the thinker language backbone and text head are quantized; audio modules, routers and Talker remain floating-point. Speculative MTP decoding and streaming are not implemented. Reference speaker extraction uses the bundled native MLX CAMPPlus encoder. Its weights retain FP32 precision. No ONNX packages are required.

from mlx_audio.sts import load
model = load('mlx-community/Ming-3.0-audio-tiny-8bit', strict=True)
result = model.generate("speech.wav", task="audio_understanding",
                        text="Describe what you hear.")
print(result.text)

Offline inference on Apple Silicon using native MLX code. The tokenizer is loaded without executing remote code.

Source checkpoint revision: 7cbe4a75c979f6ea48f1ee31998eb02bd7012233.

Quantized from the BF16 MLX export using 8-bit affine quantization, group size 64. The model-specific quantization policy preserves acoustic modules and MoE routers in floating-point precision. The native CAMPPlus speaker encoder and its Apache-2.0 license are bundled; speaker weights retain FP32 precision. The full BF16 Talker, text tokenizer and audio VAE are included under talker/. All necessary weights and tokenizers are included. No separate model downloads or ONNX packages are required. Use an MLX Audio version with Ming Tiny support.

The example repository name above may be replaced with the local export directory or your uploaded repository ID.

Validation: strictly reloaded from a relocated directory with an empty Hugging Face cache, network access blocked and ONNX packages absent. See validation.json for the tested tasks and weight inventory. SHA256SUMS covers all package files.