MASHBOX Labs

maya1 and the SNAC 24 kHz decoder

Weights only: the mashbox package, release 0.1.10, has no mashbox run command for this model. To make speech, run the original weights with Maya Research's Transformers example, which needs a CUDA GPU.

GitHub X

Maya Research's maya1 text-to-speech model as one BF16 GGUF file, with the SNAC 24 kHz decoder it speaks through as NumPy arrays. maya1 reads a voice description and the text to say, then writes SNAC audio codes; the SNAC decoder makes 24 kHz audio from those codes. Upstream lists 17 emotion tags, such as <laugh> and <whisper>, that go inline in the text.

Files

File What Size Tensors From upstream
maya1-BF16.gguf maya1, llama architecture, 28 blocks 6.60 GB 255, BF16 (norms and rope factors F32) maya-research/maya1 safetensors
snac/snac_24khz_decoder.npz SNAC quantizer codebooks and decoder 52.5 MB 104 arrays, float32 hubertsiuzdak/snac_24khz pytorch_model.bin
snac/LICENSE SNAC's licence file hubertsiuzdak/snac

The GGUF holds the weights and model settings. It has no tokenizer, so llama.cpp cannot load it on its own: the tokenizer, chat template, voice description format (prompt.txt) and emotion list (emotions.txt) are in maya-research/maya1.

Load

Both files open with standard Python readers:

pip install gguf huggingface_hub numpy
import numpy as np
from gguf import GGUFReader
from huggingface_hub import hf_hub_download

lm = GGUFReader(hf_hub_download("mashbox-labs/maya1-TPU", "maya1-BF16.gguf"))
print(len(lm.tensors), lm.tensors[0].name)   # 255 token_embd.weight

snac = np.load(hf_hub_download("mashbox-labs/maya1-TPU", "snac/snac_24khz_decoder.npz"))
print(snac["decoder.model.7.weight"].shape)  # (1, 64, 7)

Links

Original model maya-research/maya1 by Maya Research
Audio codec hubertsiuzdak/snac_24khz by Hubert Siuzdak
Companies hello@mashbox.gg

Licence

maya1: Apache-2.0, as stated by maya-research/maya1 in its model card. That repo has no separate licence file.

SNAC decoder: MIT, as stated by hubertsiuzdak/snac_24khz. snac/LICENSE is the licence file of the SNAC code repo, hubertsiuzdak/snac, unchanged.

Downloads last month
48
GGUF
Model size
3B params
Architecture
llama
Hardware compatibility
Log In to add your hardware

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Collection including mashbox-labs/maya1-TPU