A Core ML re-export of Qwen3-TTS-12Hz-0.6B-Base and the decoder of
Qwen3-TTS-Tokenizer-12Hz (both Apache-2.0, by the Qwen team). It runs
with MLComputeUnits.cpuAndNeuralEngine, without the GPU, so it keeps working in an iOS app playing in the background.
The two models that run at every frame, CodeDecoder and MultiCodeDecoder, have int8 weights (symmetric, per output
channel, from coremltools.optimize.coreml.linear_quantize_weights); every other weight, and every activation, is fp16.
MLStates. Each is read once and written once, as the program's last operation.
Programs that read a state after writing it fail to plan with error -14 without the GPU (macOS 27).[1, C, 1, T] layout, every projection a 1×1 convolution, RMSNorm written as an L2 normalization with
the weights folded into the next convolution.| Model | Inputs → outputs | Size |
|---|---|---|
TextProjector |
text token → embedding [1024,1,1] |
297 MB |
CodeEmbedder |
first codebook token → embedding | 6 MB |
CodeDecoder (prefill, decode) |
prompt or one frame → logits, hidden state | 425 MB |
MultiCodeDecoder |
hidden state, first code → codebooks 1–15 and their embedding sum | 137 MB |
SpeechDecoder |
[1,16,125] codes → 24 kHz audio |
233 MB |
Each model is present as .mlpackage and as .mlmodelc (compiled for iOS 26). vocab.json and merges.txt are the
upstream BPE tokenizer; tts_{pad,bos,eos}_embed.npy are the special text embeddings.
.cpuAndNeuralEngine)| Stage | Neural Engine share of cost |
|---|---|
CodeDecoder prefill and decode |
100% |
MultiCodeDecoder |
100% |
SpeechDecoder |
99.1% (the codebook gather runs on the CPU) |
TextProjector, CodeEmbedder |
0% (lookups, about 0.1 ms each) |
decode
and 9.7 ms in MultiCodeDecoder, against 7.4 and 16.3 ms with fp16 weights.CodeDecoder against PyTorch, teacher-forced on the same frames: 95% top-1 agreement of the logits, median KL
divergence 2.3e-3 (100% and 4.6e-4 with fp16 weights).Speaker embeddings (x-vectors) of LibriVox recordings, in the public domain:
| Voice | Language | Source |
|---|---|---|
thiago |
pt-BR | Dom Casmurro ch. LXI, read by Thiago |
peter |
en | Alice's Adventures in Wonderland ch. I, read by Peter Yearsley |
config.json lists the source URL, time range, transcript and checksum of each.
Apache-2.0, as the upstream model and tokenizer.
CodeDecoder.mlmodelc/weights/weight.bin 444 MBCodeDecoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin 444 MBTextProjector.mlmodelc/weights/weight.bin 311 MBTextProjector.mlpackage/Data/com.apple.CoreML/weights/weight.bin 311 MBSpeechDecoder.mlmodelc/weights/weight.bin 244 MBSpeechDecoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin 244 MBMultiCodeDecoder.mlmodelc/weights/weight.bin 142 MBMultiCodeDecoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin 142 MBCodeEmbedder.mlmodelc/weights/weight.bin 6 MBCodeEmbedder.mlpackage/Data/com.apple.CoreML/weights/weight.bin 6 MBCodeDecoder.mlmodelc/coremldata.bin 0.6 kBMultiCodeDecoder.mlmodelc/coremldata.bin 0.5 kBCodeEmbedder.mlmodelc/coremldata.bin 0.4 kBTextProjector.mlmodelc/coremldata.bin 0.4 kBSpeechDecoder.mlmodelc/coremldata.bin 0.4 kBCodeDecoder.mlmodelc/analytics/coremldata.bin 0.2 kBCodeEmbedder.mlmodelc/analytics/coremldata.bin 0.2 kBMultiCodeDecoder.mlmodelc/analytics/coremldata.bin 0.2 kBSpeechDecoder.mlmodelc/analytics/coremldata.bin 0.2 kBTextProjector.mlmodelc/analytics/coremldata.bin 0.2 kBHave the files? Submit a torrent.
Submit a torrentHave the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your settings, and read the script before running it.
curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo itsmechico/qwen3-tts-0.6b-base-coreml-ane ./model-folderQwen/Qwen3-TTS-12Hz-0.6B-Base