Notices and attribution

This repository is an independent MLX (Apple Silicon) port of Maya 2 Native by Maya Research. It is not an official Maya Research release.

Maya 2 Native (model weights and code), Apache-2.0

Model: maya-research/maya2-native, © Maya Research, Apache-2.0 (LICENSE, copied unchanged from the release).
Code: MayaResearch/maya2-native, commit c6b3768, Apache-2.0.
Changes in this repository:
- The weights were re-laid out for MLX.
- models/maya2-8bit quantizes most backbone tensors, the token embeddings and the feedback table to 8 bits (the recipe is in its quant.json).
- The inference code was rewritten for MLX in maya2_mlx/.
- maya3/guards.py, maya3/assets/, maya3/text/, maya3/__init__.py and configs/text_frames_prior.json are copied unchanged from the official repository.
Built-in voice names identify synthetic personas, not real people. Follow the original model card's intended-use and responsible-deployment section: disclose synthetic speech, no impersonation or deception.

Kyutai Mimi (audio codec weights), CC-BY-4.0

models/*/mimi.safetensors holds Mimi's decoder: the quantizer codebooks, the upsampler, the decoder transformer and the SEANet decoder.
Changes: convolution weights re-laid out for MLX, codebook embeddings stored as computed (embed_sum / cluster usage), RoPE tables precomputed, encoder weights left out. Values are otherwise unchanged (fp32).

Qwen3 (backbone architecture), Apache-2.0

The backbone has the Qwen3-1.7B architecture (Qwen team, Alibaba, Apache-2.0). Its weights here are Maya 2 Native's trained weights, not Qwen's.

Citation of the original model

@misc{maya2native2026,
  title        = {Maya 2 Native: Streaming Multilingual Text-to-Speech},
  author       = {{Maya Research}},
  year         = {2026},
  url          = {https://huggingface.co/maya-research/maya2-native},
  note         = {Release 1.0}
}