Model reference · open weights
Ming-3.0-tts-ONNX is an open-weight audio or speech model from onnx-community.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the onnx-community/Ming-3.0-tts-0.5B-ONNX model card, 2026-10-29
What it is
| Released by | onnx-community |
|---|---|
| Released | 2026-10-27 |
From the model card
BF16 component ONNX export of Ming-3.0-tts-0.5B. Ten graphs are exported from the source neural components at Opset 18 / IR 10; the upstream CAMPPlus speaker graph is reused with its original Opset 14 / IR 7. Source FP32 buffers are retained.
| Graph | Purpose |
|---|---|
token_embedding.onnx | Token IDs to embeddings |
llm_prefill.onnx, llm_decode.onnx | Hidden states and explicit text-backbone KV cache |
speaker_embedding.onnx | Upstream speaker feature graph |
speaker_projection.onnx | Speaker embedding projection |
stop_head.onnx | Stop logits |
aggregator.onnx | Acoustic latent patch to next-step embedding |
dit_cfg_step.onnx | Conditional and unconditional flow-field predictions |
audio_encoder.onnx | Reference waveform and explicit posterior noise to latents |
audio_decoder_spectrum.onnx | Full latent sequence to real/imaginary spectrum |
waveform_reconstruct.onnx | Spectrum to waveform |
All graphs are under onnx/. This component package requires host-side prompt
construction, audio resampling/features, flow integration, stopping logic and cache
orchestration. CAMPPlus expects prepared features, not arbitrary raw waveform input.
Batch size is one and audio reconstruction is nonstreaming. No Ming FP8 artifact
is included. Exact tensor contracts are in model_interface.json.
Audio backbones use explicit causal/sliding masks and per-instance layer-count configuration. Complex inverse FFT is represented by real inverse DFT matrix products and overlap-add normalization; computation order is not bitwise identical. BF16 trigonometric operations use FP32 computation and BF16 rounding boundaries. Clipping bounds are explicit tensors on the input device.
AudioEncoder capture used draft_export, with 26 reported
shape-guard/soundness issues. Symbolic dimensions remain in the graph, but dynamic
length correctness has not been established by runtime tests. In the current
contract, integer audio blocks range from 2 to 512; padded waveform length is
3528 times the block count, with the actual length supplied separately.
Export and static ONNX structural/type/shape checks have completed. Runtime execution,
numerical alignment, quality and performance benchmarks have not been performed for
these artifacts. The status is EXPORTED_NOT_VALIDATED.
Configuration and tokenizer resources are at the repository root. Graphs are under
onnx/; their shared external data is onnx/weights/model.bin. Keep this layout
when moving the model. model_interface.json describes names, dtypes, shapes and
state mappings. Files absent from the upstream checkpoint are not synthesized.
conversion/ contains only this model's export code, a static format checker,
dependencies and English reproduction instructions. The export report is
conversion/export_report.json. Original source checkpoints and trusted identity
manifests are separate reproduction inputs.
python conversion/scripts/check_onnx_format.py --model-dir .
# Optional: also stream and verify all recorded file/weight hashes.
python conversion/scripts/check_onnx_format.py --model-dir . --verify-hashes
These commands do not run inference. For export commands and environment details, see conversion/README.md.
See the original license and/or LICENSE_NOTICE.md at the repository root for
upstream declarations and source references.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.