AEmotionStudio/ming-tts-models

Ming-3.0-tts — mirror for Mæstræa

A verbatim mirror of inclusionAI's Ming-3.0-tts checkpoints, laid out for the Mæstræa AI Workstation (AI ▸ Voice ▸ Ming TTS). All credit for the models goes to inclusionAI: see the original repository and the model cards for 0.5B and tiny.

License

MIT, © 2025 inclusionAI — the same license as the upstream code and weights (see LICENSE). Nothing here is relicensed.

Upstream's usage note applies: use voice cloning and built-in voices only for research and legitimate creative purposes, do not impersonate real people without consent, and label synthetic speech where it could be mistaken for a real recording.

Layout

Path	Upstream	Contents
0.5b/	inclusionAI/Ming-3.0-tts-0.5B @ 8d66b70a	Qwen2.5-0.5B backbone + DiT flow head + audio VAE (model.safetensors, bf16), CAM++ speaker encoder (campplus.onnx), tokenizer, upstream model card
0.5b/tn/	GitHub inclusionAI/Ming-3.0-tts @ 028c1eac (frontend/talker_tn/)	The four Chinese/English text-normalization FST grammars the 0.5B pipeline loads
tiny/	inclusionAI/Ming-3.0-tts-tiny @ 7bd9446a	Ling-V3 MoE (7.9B total, 1.3B active) in 30 bf16 shards + index, CAM++, tokenizer, chat template, upstream model card

Every file is byte-identical to its source. <lane>/sha256.json records the digest, size and source of each file. The upstream remote-code .py files are not mirrored: Mæstræa vendors the inference code and never loads remote code.

What changed

Nothing in the weights. Mæstræa's runtime differs from upstream's only in how it loads them. It runs on newer transformers, with a decode-mask fix, CUDA-graphed sampling heads, and a GPU expert cache that pages tiny's routed experts on 12 GB cards. Its built-in voice list is limited to the original studio voices; the upstream catalogue's character and actor voices are not offered.