---
title: Hojo-TTS-Light-80M-INT8
canonical_url: "https://www.modelscope.cn/models/nocmtone/Hojo-TTS-Light-80M-INT8"
md_url: "https://www.modelscope.cn/models/nocmtone/Hojo-TTS-Light-80M-INT8.md"
repository: nocmtone/Hojo-TTS-Light-80M-INT8
chinese_name: Hojo-TTS-Light-80M-INT8
last_updated: 2026-09-17
license: apache-2.0
pipeline_tag: text-to-speech
tasks:
  - text-to-speech
model_type:
  - qwen3
architectures:
  - Qwen3ForCausalLM
base_model:
  - HojoAI/Hojo-TTS-Light
base_model_relation: finetune
library_name:
  - onnx
downloads: 0
stars: 0
tags:
  - tts
  - voice-cloning
  - onnx
  - int8
  - quantization
---

# Hojo-TTS-Light-80M-INT8

> Hojo-TTS-Light-80M-INT8 - nocmtone 在 ModelScope 开源的模型。Hojo-TTS-Light-80M-INT8

nocmtone/Hojo-TTS-Light-80M-INT8 是 ModelScope 魔搭社区上的text-to-speech模型，采用 apache-2.0 许可，基于 HojoAI/Hojo-TTS-Light 构建。

- **Repository**: nocmtone/Hojo-TTS-Light-80M-INT8
- **License**: apache-2.0
- **Tasks**: text-to-speech
- **Base model**: HojoAI/Hojo-TTS-Light
- **Tags**: tts, voice-cloning, onnx, int8, quantization
- **Downloads**: 0
- **Stars**: 0
- **Last updated**: 2026-09-17

Source: https://www.modelscope.cn/models/nocmtone/Hojo-TTS-Light-80M-INT8

---

# Hojo-TTS-Light-80M-INT8

int8 quantization of the decoder LM (`Hojo-TTS-Light-llm.onnx`) from
[HojoAI/Hojo-TTS-Light](https://huggingface.co/HojoAI/Hojo-TTS-Light) at revision
`d9c02af95496b6363b44944326d71491ab8905b0`.

The 80M checkpoint is the voice-cloning model: audio is synthesized from a
speaker vector and reference codes extracted from a reference clip, rather than
from a preset voice. This repository mirrors the snapshot verbatim and replaces
exactly one graph.

## Why only the LM

Synthesis is autoregressive: the LM runs once per generated audio token, so it
dominates the cost of a sentence. The encoder, speaker encoder and vocoder graphs
are deliberately left at their original precision — dynamic weight-only
quantization buys little on their convolutions, and they are where audible
artifacts would come from.

A useful consequence is that the quantized and full-precision variants produce a
**bit-identical** speaker vector and reference codes from the same reference clip
(measured: norm `275.4591`, 146 codes in both). Only the LM output differs, which
makes an A/B comparison of the two variants meaningful.

## Files

Every file below is part of the installer's snapshot contract — file names are
case sensitive and the byte counts and hashes are verified at download time.
`Hojo-TTS-Light-llm.onnx` is the only quantized graph; the rest are byte-identical
to the upstream snapshot.

| file | bytes | sha256 |
|---|---|---|
| `Hojo-TTS-Light-decoder.onnx` | 151719689 | `bae05e186dc4e1b37ba354efb4cc0a160ae48661f574cf558c77dc3e09d30d81` |
| `Hojo-TTS-Light-encoder.onnx` | 82140638 | `c897b928dee1ebd95309a114a73765a867aa916d98fe2151811e0f0cdde95dab` |
| **`Hojo-TTS-Light-llm.onnx`** | **82237257** | `61dbc211dac1729dc07b2c7311ff135605f5371e7eb01db78c7a3dbeb637365a` |
| `Hojo-TTS-Light-speaker.onnx` | 24160472 | `2297f00fe783e6f985cf4cfab333f313e53929c36f7348d5617c11c8acc3f963` |
| `Hojo-TTS-Light-voice.npz` | 36182300 | `82be65c9687f03704cb2941c8066727165c0800a18e1d2b816afc60df20b45a8` |
| `config.json` | 1222 | `f0ec939c99a1d053c483a806db9ff9f776cec87ddb279c9ae94822004861c505` |
| `tokenizer.json` | 2071007 | `395daae9cd5a4e166f1892d22e01c28a22421cfc92bc97aeb5d2b0fc6eaac499` |
| `tokenizer_config.json` | 455 | `e2df8b54623ab7f105c0ddfa4fd84e0e50353b2265550d9eb6ed4cd521d41d13` |

The upstream LM is BF16 (161483978 bytes) and is promoted to FP32 (322182641
bytes) at install time. This file replaces that intermediate step: it is already
int8, so the promotion pass finds no BF16 tensors and leaves the numerics
untouched (verified: identical initializer count and dtypes, and bit-identical
step outputs).

`token_embedding.f32` is deliberately **not** stored here — it is generated at
install time from `voice.npz`.

## How it was produced

```bash
pip install onnxruntime
python tool/quantize_hojo_tts_llm.py <installed-model-dir> --verify
```

which runs:

```python
onnxruntime.quantization.quantize_dynamic(
    source, target,
    weight_type=QuantType.QInt8,
    op_types_to_quantize=["MatMul", "Gemm"],
    per_channel=True,
    reduce_range=False,
)
```

## Measured

On an i5-12600KF (16 logical cores, 6P+4E). One decode step of the LM graph:

| | 4 threads | 8 threads | size |
|---|---|---|---|
| fp32 | 41.7 ms | 45.0 ms | 322.2 MB |
| **int8** | **16.1 ms** | **14.1 ms** | **82.2 MB** |

End to end through the real clone pipeline (encode a reference clip, then read a
Chinese sentence with the cloned voice), real-time factor per utterance:

| | fp32 | int8 |
|---|---|---|
| short sentence | 2.475 / 2.479 / 3.353 | **1.014 / 1.018 / 1.245** |
| longer sentence | 2.600 | **1.030** |

Roughly 2.5x, which moves this model from clearly slower than real time to
approximately real time. The end-to-end gain is smaller than the per-step gain
because the encoder and vocoder — untouched by quantization — make up a larger
share of the total here than they do in the 40M model.

Quantization also shifts the optimal thread count for this graph: 8 threads beat
4 for int8, while the reverse holds for fp32.

## License

Apache-2.0, inherited from the base model (see `LICENSE`). This repository is a
precision-only derivative of the upstream weights; the architecture is unchanged.