idefinity/webfm-ming-3.0-tts

WebFM Ming-3.0-tts 0.5B

Experimental WebFM assets for replacing F5-TTS with inclusionAI/Ming-3.0-tts-0.5B.

What is here

• distill/dit4.onnx: 4-step CFG-free flow DiT student (fp32 ONNX).
• distill/dit4_int4.onnx: student DiT with ONNX Runtime MatMulNBits (4-bit weights, block size 32, symmetric, accuracy level 4).
• distill/student.pt: PyTorch state dict for the student DiT.
• samples/: before/after listening examples.
• webfm-ming-tts-int4.zip: the full experimental int4 ONNX bundle (LM prefill/decode, DiT, aggregator, encoder, vocoder, host.npz, manifest, samples). ~1.1 GB.

Listening examples

Text: 雨落在窗台上，远处的灯一盏一盏亮起来。

Stage	File	Duration
Teacher, 10-step CFG=2	teacher_10step.wav	3.70 s
Student, 4-step CFG-free	student_4step.wav	3.68 s
Student + int4 DiT	student_4step_dit_int4.wav	3.68 s

The int4 sample uses the PyTorch LM, aggregator and vocoder with the quantized ONNX DiT. It is a flow-head quantization example, not a full int4-LM sample.

Distillation

• Teacher: 10 Euler steps, CFG=2, fp32 on Tesla T4.
• Student: 4 Euler steps, CFG-free, initialized from the teacher DiT.
• Data: 200 SFW Chinese utterances, 13,112 latent frames, generated from books/story.txt with the lingguang prompt voice.
• Training: 2,000 optimizer steps, batch 16, AdamW lr 1e-4, fp32.
• Loss: 1.0234 -> 0.2379.
• Student ONNX parity: max_abs=1.19e-6 against PyTorch.
• Int4 DiT probe: cosine 0.9961, max_abs=0.5156.

Int4 bundle warning

The bundled LM is quantized to 4-bit MatMulNBits but its validation is much worse than the DiT:

• lm_prefill: hidden cosine 0.881, K/V cosine 0.999 / 0.925.
• lm_decode: hidden cosine 0.555, K/V cosine 0.996 / 0.705.
• dit4: cosine 0.996; aggregator: 0.999; enc: 0.99996; voc: 0.998.

The LM is the sensitive component. Treat the int4 LM tier as experimental. A safer shipping tier is likely fp16 or int8 LM with 4-bit DiT/vocoder.

Status

Research output, SFW only. The base model is MIT licensed and its usage disclaimer applies. Full WebGPU validation and full int4-LM listening evaluation are still pending.