README.md

---
library_name: mlx
pipeline_tag: text-to-speech
base_model: IndexTeam/IndexTTS-2.5
license: other
license_name: bilibili-model-use-license-agreement
license_link: https://huggingface.co/IndexTeam/IndexTTS-2.5/blob/d0aa86e75bb6f3437f3831e95056fa72842d89ef/LICENSE
language:
- zh
- en
- ja
- es
- ar
tags:
- indextts2
- indextts-2.5
- mlx
- apple-silicon
- text-to-speech
- voice-cloning
- emotion-control
- 8bit
---

# mlx-indextts2-2.5-8bit

IndexTTS 2.5 converted to native MLX weights for Apple Silicon. Use it with
[`vanch007/mlx-indextts2`](https://github.com/vanch007/mlx-indextts2).

## Variant

- Precision / quantization: **8bit**
- Approximate local size: **1.6GB**
- Source model: [`IndexTeam/IndexTTS-2.5`](https://huggingface.co/IndexTeam/IndexTTS-2.5)
- Source revision: `d0aa86e75bb6f3437f3831e95056fa72842d89ef`
- Conversion status: strict manifest **pass**
- Details: GPT uses persistent 8-bit group quantization; the remaining components use float16. This is the recommended default variant.

The model supports Chinese, English, Japanese, Spanish, and Arabic, including
cross-lingual voice transfer, separate speaker/emotion references, manual and
text-derived emotion control, and Pinyin/CMU/Kana pronunciation annotations.

## Download and Run

```bash
git clone https://github.com/vanch007/mlx-indextts2.git
cd mlx-indextts2
uv sync --extra v25

hf download vanch007/mlx-indextts2-2.5-8bit \
  --local-dir models/mlx-indextts2-2.5-8bit

uv run mlx-indextts generate \
  --profile v25 \
  -m models/mlx-indextts2-2.5-8bit \
  --language zh \
  -r /path/to/reference.wav \
  -t "你好，这是 IndexTTS 2.5 的 MLX 推理测试。" \
  -o output.wav
```

See the [project README](https://github.com/vanch007/mlx-indextts2#quick-start) and
[IndexTTS 2.5 guide](https://github.com/vanch007/mlx-indextts2/blob/main/docs/indextts-2.5.md) for batch,
API, WebUI, streaming, duration, pronunciation, and emotion-control examples.

## Included Files

- `gpt.safetensors`
- `codec.safetensors`
- `s2mel.safetensors`
- `bigvgan.safetensors`
- `multilingual_zh_ja_yue_char_del.tiktoken`
- `config.yaml` and `config.json`
- `model_manifest.json` and `conversion_report.json`
- reference preprocessing statistics/features
- `LICENSE`

The runtime additionally resolves the manifest-declared
`facebook/w2v-bert-2.0` and `funasr/campplus` preprocessing dependencies.

## Validation Boundary

The converted component coverage and strict load checks passed. The project
validation matrix also covers five-language synthesis, ASR sanity checks,
speaker similarity, emotion modes, cross-lingual generation, batch/API/WebUI,
and completed-segment streaming. Results are hardware- and sample-specific;
human listening remains recommended before production use.

## License and Derivative Notice

The repository code is MIT-licensed, but these converted model weights are a
Derivative Work governed by the included **bilibili Model Use License
Agreement**. Review its commercial thresholds, downstream obligations, use
restrictions, and prohibited high-risk scenarios before downloading or using
the model.

Any modifications made to the original model in this Derivative Work are not
endorsed, warranted, or guaranteed by the original right-holder of the
original model, and the original right-holder disclaims all liability related
to this Derivative Work.