Harsha001/Maya2-Native-GGUF

Maya2 Native GGUF

GGUF files of maya-research/maya2-native, a multilingual Indic text-to-speech model, for the ROCmFPX fork of llama.cpp. The runtime is native C++ (ggml): no Python, no PyTorch. It works on CPU, Vulkan (AMD Strix Halo and other GPUs) and CUDA.

Files

File	Size	What it is
Maya2-Native-BF16.gguf	3.6 GB	backbone, full precision
Maya2-Native-Q8_0.gguf	1.9 GB	backbone, 8-bit (recommended)
Maya2-Native-Q4_K_M.gguf	1.1 GB	backbone, 4-bit
mmproj-Maya2-Native-BF16.gguf	1.5 GB	depth decoder + Mimi audio decoder, full precision
mmproj-Maya2-Native-Q8_0.gguf	0.8 GB	depth decoder + Mimi audio decoder, 8-bit (recommended)

You need one backbone file and one mmproj file.

Run

Build ROCmFPX (Vulkan example):

git clone https://github.com/HarshaVardhan3002/ROCmFPX && cd ROCmFPX
cmake -B build -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --target llama-tts -j

Speak:

build/bin/llama-tts -m Maya2-Native-Q8_0.gguf -mm mmproj-Maya2-Native-Q8_0.gguf -ngl 99 \
    --tts-lang hi -p "{Vaidehi}: नमस्ते, आज मौसम बहुत अच्छा है।" --output out.wav

• Languages (--tts-lang): en hi bn ta te kn ml mr gu pa or as
• Voices (start the prompt with {Name}:): Ahana, Vaidehi, Malini, Charvi, Jahnavi, Anirban, Nakul, Vasant, Siddhant, Satvik, Abhay, Nivedita, Ishwari, Ritvik, Vrinda, Sreekar, Kayal, Shaunak, Akshat, Tanushree. Default: Ahana.
• -s <seed> gives the same audio every time.

How close is it to the original?

Checked against the PyTorch reference with the same codes fed in (teacher forcing), English test sentence, 54 frames:

Backbone / mmproj	backbone logits cos	depth logits cos	audio cos	top choice matches
F32 / F32	0.9999998	0.9999999	0.9999999	55 / 55
BF16 / BF16	0.999997	0.99999	0.9999995	55 / 55
Q8_0 / Q8_0	0.9997	0.9995	0.99999	54 / 55
Q4_K_M / Q8_0	0.986	0.994	0.99999	50 / 55

Whisper transcribes the free-running output of every quant word for word in English. Hindi output was checked with three seeds.

Speed on an NVIDIA GB10 (CUDA, Q4_K_M): about 40 frames per second, 3.2 times faster than real time (the model makes 12.5 frames per second of audio). Strix Halo Vulkan numbers come from the community, please share yours.

How it works

• Backbone (Qwen3, 28 layers): reads the text and predicts code 0 of each audio frame
• Depth decoder (44 small transformer layers in 11 groups): predicts codes 1 to 31 of the frame
• Mimi (kyutai): turns the 32 codes per frame into 24 kHz audio

The sampling settings and the stop rule follow the reference engine.

Credits and licenses

• Model: maya-research/maya2-native, Apache 2.0
• Audio decoder: kyutai/mimi, CC-BY 4.0
• Runtime: ROCmFPX, built on llama.cpp (MIT)