VoxCPM2-GGUF

VoxCPM2 is an open-source, 2-billion-parameter tokenizer-free diffusion autoregressive Text-to-Speech (TTS) model built on the MiniCPM-4 backbone and trained on over 2 million hours of speech data. Supporting 30 languages and 9 Chinese dialects without requiring explicit language tags, it utilizes an asymmetric AudioVAE V2 to generate 48kHz studio-quality audio from 16kHz inputs. Its standout capabilities include Voice Design (creating novel voices solely from natural-language descriptions), Controllable Cloning (cloning voices from short clips with style and emotion steering), and Ultimate Cloning (high-fidelity audio continuation using a reference clip and its transcript). Requiring roughly 8 GB of VRAM, it supports real-time streaming (reaching a Real-Time Factor as low as ~0.13 with Nano-vLLM), offers LoRA and full fine-tuning with as little as 5–10 minutes of data, and is released under a commercially permissive Apache-2.0 license.

Model Files

File Name Quant Type File Size File Link Description
VoxCPM2.BF16.gguf BF16 2.95 GB Link Full BF16 weights. Highest quality, largest file size.
VoxCPM2.Q3_K_L.gguf Q3_K_L 818 MB Link Lower quality but usable, good for low RAM availability.
VoxCPM2.Q3_K_M.gguf Q3_K_M 759 MB Link Low quality, smallest file size.
VoxCPM2.Q4_K_M.gguf Q4_K_M 916 MB Link Good quality, default size for most use cases, recommended.
VoxCPM2.Q4_K_S.gguf Q4_K_S 874 MB Link Slightly lower quality with more space savings, recommended.
VoxCPM2.Q5_K_M.gguf Q5_K_M 1.06 GB Link High quality, recommended.
VoxCPM2.Q5_K_S.gguf Q5_K_S 1.03 GB Link High quality with slight space savings, recommended.
VoxCPM2.Q6_K.gguf Q6_K 1.21 GB Link Very high quality, near perfect, recommended.
VoxCPM2.Q8_0.gguf Q8_0 1.57 GB Link Extremely high quality, generally unneeded but max available standard quant.
VoxCPM2.mmproj-bf16.gguf mmproj-bf16 2.01 GB Link Multimodal projection file in BF16 format. Used for audio/voice capabilities.

Run with llama.cpp

Install dependencies and clone llama.cpp

!pip install transformers huggingface-hub

!git clone https://github.com/ggml-org/llama.cpp.git
%cd /data/llama.cpp

Download GGUF model files

Download the BF16 model and BF16 multimodal projector:

from huggingface_hub import snapshot_download

# Target local directory
local_dir = "./model/VoxCPM2-GGUF"

# Download only the specified files
snapshot_download(
    repo_id="prithivMLmods/VoxCPM2-GGUF",
    local_dir=local_dir,
    allow_patterns=[
        "VoxCPM2.BF16.gguf",
        "VoxCPM2.mmproj-bf16.gguf",
    ],
    resume_download=True,
)

Build llama-tts

Build the TTS CLI with the CUDA backend enabled:

!cmake --build . --config Release -j$(nproc) --target llama-tts

Text-to-speech generation (GPU offloading)

Run inference using llama-tts with GPU offloading enabled:

!/data/llama.cpp/build/bin/llama-tts \
    -m /data/llama.cpp/model/VoxCPM2-GGUF/VoxCPM2.BF16.gguf \
    -mm /data/llama.cpp/model/VoxCPM2-GGUF/VoxCPM2.mmproj-bf16.gguf \
    -p "VoxCPM2 is a tokenizer-free, diffusion autoregressive Text-to-Speech model with 2 billion parameters, support for 30 languages, and 48 kHz audio output. It is trained on over 2 million hours of multilingual speech data." \
    -ngl 99 \
    -n 1000 \
    --seed 42 \
    --output /data/llama.cpp/model/VoxCPM2-GGUF/out.wav

Output

The generated audio file will be saved at:

/data/llama.cpp/model/VoxCPM2-GGUF/out.wav

llama.cpp

LLM inference in C/C++ β€” https://github.com/ggml-org/llama.cpp

Quantized using llama.cpp by prithivMLmods.

Downloads last month
18
GGUF
Model size
1B params
Architecture
minicpm
Hardware compatibility
Log In to add your hardware

3-bit

4-bit

5-bit

6-bit

8-bit

16-bit

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

Model tree for prithivMLmods/VoxCPM2-GGUF

Base model

openbmb/VoxCPM2
Quantized
(29)
this model

Collections including prithivMLmods/VoxCPM2-GGUF