---
license: other
license_name: boson-higgs-audio-v3-research-and-non-commercial-license
license_link: LICENSE
base_model:
  - bosonai/higgs-audio-v3-tts-4b
base_model_relation: quantized
library_name: transformers
pipeline_tag: text-to-speech
tags:
  - text-to-speech
  - speech-generation
  - higgs-audio
  - qwen3
  - quantization
  - 4-bit
  - mlx
---

# Higgs-Audio-v3-TTS-4bit-MLX

4-bit MLX 4-bit artifact for [`bosonai/higgs-audio-v3-tts-4b`](https://huggingface.co/bosonai/higgs-audio-v3-tts-4b).

## Scope

This is a transformer-body quantized artifact, not a complete drop-in runtime yet. It quantizes `body.layers.*` attention/MLP 2D weights and preserves the Higgs audio tokenizer/vocoder, fused modality embedding/head, norms, biases, and non-2D tensors.

Higgs Audio v3 TTS uses a custom `HiggsMultimodalQwen3ForConditionalGeneration` architecture with 8 audio codebooks, delayed multi-codebook generation, and waveform decode. Current vanilla Transformers in the tested environment does not instantiate this architecture, so runtime integration must be done through SGLang-Omni or a custom loader.

## Quantization Report

- Quantized tensors: `252`
- Quantized parameter fraction seen: `0.7805`
- Mean relative L2: `0.066956`
- Max relative L2: `0.103319`
- Max absolute error: `0.075130`

See:

- `quantization_config.json`
- `quant_error_report.json`
- `tensor_manifest.json`

## License

Released under the upstream Boson Higgs Audio v3 research and non-commercial license. Production, hosted APIs, or revenue-generating use requires a separate commercial license from Boson AI.