MioTTS-0.6B-q8-webgpu

MioTTS-0.6B, int8 for the browser

Aratako/MioTTS-0.6B quantized to per-row int8 so it can run in a web browser on WebGPU, as part of voxshot.

This is a repackaging, not a new model. Every weight comes from the Apache-2.0 checkpoint above; nothing was trained, fine-tuned or distilled. Please cite and credit the original.

Why this exists

A browser cannot afford the bf16 checkpoint, and the op library this runs on (web-xpu-ops) has exactly one quantized kernel: a fused W8A32 GEMV over per-row absmax int8. So that is the format. There is no f16 path — no shader-f16 kernels exist upstream.

Files

manifest.json is the index: every tensor entry carries rows, cols and a byte offset into the file it lives in. Nothing else is needed to read these.

The quantization

Per-row absmax symmetric int8. For each output row, scale = max|w| / 127 and code = round(w / scale) clamped to [-127, 127]; an all-zero row gets scale = 1. Rounding is half-up, matched bit-for-bit against the consumer's own quantize op on real rows of this checkpoint.

Quantized: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj per layer, and embed_tokens once — the embeddings are tied, so that one matrix serves as both the lookup table and the output projection. Left in f32: the layer norms and the per-head q/k-norm gains, 291 KB in total, where quantization would buy nothing and cost accuracy.

What it costs, measured

Against the f32 original, stage by stage:

• element error at the theoretical bound, ~scale/2 ≈ 0.0039 of each row's peak
• prompt logits within 4.8e-2 peak-relative (Japanese) and 1.2e-2 (English), with the argmax unchanged at the prompt boundary in both
• greedy generation diverges from f32 at step 6 on a Japanese test sentence: an adjacent-codebook near-tie flips, and the trajectory then differs while remaining well-formed — 87 speech tokens and a clean EOS against f32's 83.

That last point matters if you plan to compare outputs: int8 does not reproduce the f32 token sequence, and it is not supposed to. Compare against an int8 reference, or compare the audio.

RoPE channels are permuted

q_proj and k_proj rows, and the q_norm / k_norm gains, are relabeled from Hugging Face's rotate-half pairing (j, j+64) to the adjacent-pair pairing (2i, 2i+1) that the consumer's RoPE kernel implements. Per head: j → 2j for j < 64, else j → 2(j-64)+1.

manifest.json records this as "ropePermuted": true. The unpermuted gains are also stored, as qNormRaw / kNormRaw, so the original order is recoverable.

If you feed these weights to a stock Hugging Face implementation without undoing the permutation, you will get noise. They are shaped for one consumer.

Provenance

Produced by spike/miotts/convert_weights.py from:

Aratako/MioTTS-0.6B  model.safetensors
  1,217,825,224 bytes
  sha256 d0f85e5d9799c1589d7ba7c1c7830a7e2b6e4424ed07da2f76a56bc2ed7dc449

The converter re-hashes that file and aborts on a mismatch, and manifest.json records the digest, so these weights and the checkpoint they came from cannot drift apart unnoticed.

Architecture, for reference

Qwen3, 28 layers, hidden 1024, 16 query heads / 8 KV heads, head dim 128, FFN 3072 SwiGLU, RMSNorm eps 1e-6, RoPE theta 1e6, tied embeddings, vocab 164,480. Speech tokens are <|s_n|> = id 151669 + n for n in 0..12799, which index Aratako/MioCodec-25Hz-24kHz at 25 tokens per second.

License and attribution

Apache-2.0, inherited from the source checkpoint.

• Model: Aratako/MioTTS-0.6B by Chihiro Arata
• Codec it pairs with: Aratako/MioCodec-25Hz-24kHz (MIT)

@misc{miotts,
  author = {Chihiro Arata},
  title = {MioTTS: Lightweight and Fast LLM-based Text-to-Speech},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{https://huggingface.co/collections/Aratako/miotts}}
}