Ming-3.0 for ComfyUI
ComfyUI-ready weights of inclusionAI's Ming-3.0-tts
speech synthesis models and Ming-3.0-audio-tiny, laid out like ComfyUI's models/ folder, for the
ComfyUI-MingTTS node pack. The pack is pure ComfyUI:
the TTS models load with Load Diffusion Model, the audio VAE with Load VAE, and audio-tiny
with Load CLIP for the stock TextGenerate node.
File (= path under ComfyUI/models/) |
Size | Notes |
|---|---|---|
diffusion_models/ming-3.0-tts-0.5b_bf16.safetensors |
1.4 GB | Qwen2.5-0.5B backbone, 1.6x real time on an RTX 5090 |
diffusion_models/ming-3.0-tts-tiny_int8_convrot.safetensors |
8.7 GB | Ling V3 MoE (7.9B / 1.3B active), int8 + ConvRot, ~1x real time, fits 12 GB |
diffusion_models/ming-3.0-tts-tiny_bf16.safetensors |
15.9 GB | the unquantized tiny |
vae/ming-3.0_audio_vae.safetensors |
1.6 GB | 44.1 kHz AudioVAE + CAM++ speaker encoder, shared by all models (needed) |
text_encoders/ming-3.0-audio-tiny_int8_convrot.safetensors |
9.8 GB | speech recognition / audio understanding (optional), int8 + ConvRot |
text_encoders/ming-3.0-audio-tiny_bf16.safetensors |
17.1 GB | the unquantized audio-tiny |
How they were made
From the released checkpoints with the pack's scripts. Tensor values are unchanged apart from dtype:
scripts/convert.pysplits each checkpoint into the TTS model and the shared AudioVAE (byte-identical across all Ming-3.0 releases). It stacks the tiny's 128 MoE experts per layer into ComfyUI expert banks, and adds the CAM++ speaker encoder (read fromcampplus.onnx) to the VAE file. It drops the unused text heads and audio-tiny's multi-token-prediction layer.scripts/quantize_int8.py: int8 + ConvRot (block-Hadamard rotation, per-channel scales) for every MoE expert (tiny and audio-tiny) and the attention/MLP linears (mean weight error 0.78%). The router, the KDA decay gates, the flow head and the embeddings stay bf16/fp32.
The ComfyUI port matches the reference runtimes in fp32: the 0.5B to within 2.7e-6, the tiny to within 3e-7 per layer. Details are in the node pack's README.
License
The weights are inclusionAI's, released under the MIT License (Copyright (c) 2025 inclusionAI): Ming-3.0-tts-0.5B, Ming-3.0-tts-tiny, Ming-3.0-audio-tiny. The speaker encoder is CAM++ from 3D-Speaker (Apache 2.0).