VoxCPM2-GGUF 30-Language Speech Synthesis Model: 4GB+ VRAM, File List

VoxCPM2-GGUF 30-Language Speech Synthesis Model: 4GB+ VRAM, File List

At a Glance

Item Value
Repository ggml-org/VoxCPM2-GGUF
Publisher guide OpenBMB: models and licenses
Published 2026-11-22
License apache-2.0
Formats GGUF
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

ggml-org has released “ggml-org/VoxCPM2-GGUF", which is the GGUF format version of the text-to-speech (TTS) model “VoxCPM2" developed by OpenBMB.

This model is a text-to-speech model that takes text input and outputs natural speech. It is a converted version of the original model “openbmb/VoxCPM2" designed to be easily run via GGUF-compatible environments such as llama.cpp or command-line tools.

Specifications

The main specifications of “VoxCPM2" and this GGUF version stated in the documentation are as follows:

  • Parameters: 2B (backbone based on MiniCPM-4, 2B parameters in total)
  • Architecture: Tokenizer-free Diffusion Autoregressive (LocEnc → TSLM → RALM → LocDiT)
  • Audio VAE: AudioVAE V2 (asymmetric encode/decode, 16kHz in → 48kHz out)
  • Output Specifications: 48kHz (studio-quality audio output)
  • Supported Languages: 30 languages (Japanese, English, Chinese, Arabic, Burmese, Danish, Dutch, Finnish, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Khmer, Korean, Lao, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Tagalog, Thai, Turkish, Vietnamese) and 9 Chinese dialects (Sichuanese, Cantonese, Wu, Northeast Mandarin, Henan Mandarin, Shaanxi Mandarin, Shandong Mandarin, Tianjin Mandarin, Min Nan)
  • Training Data Size: 2M+ hours (over 2 million hours of multilingual speech data)
  • LM Token Rate: 6.25 Hz
  • Max Sequence Length: 8192 tokens
  • License Terms: Apache-2.0 (commercial use is permitted)

Performance

The model card for the base model “openbmb/VoxCPM2" provides the following detailed table regarding model structure and performance.

Property Value
Architecture Tokenizer-free Diffusion Autoregressive (LocEnc → TSLM → RALM → LocDiT)
Backbone Based on MiniCPM-4, totally 2B parameters
Audio VAE AudioVAE V2 (asymmetric encode/decode, 16kHz in → 48kHz out)
Training Data 2M+ hours multilingual speech
LM Token Rate 6.25 Hz
Max Sequence Length 8192 tokens
dtype bfloat16
VRAM ~8 GB
RTF (RTX 4090) ~0.30 (standard) / ~0.13 (Nano-vLLM)

As the above figures show, inference using an NVIDIA RTX 4090 achieves a real-time factor (RTF) of about 0.30 standard and about 0.13 with Nano-vLLM applied, providing processing performance suitable for streaming generation. On the other hand, official mentions are also made of limitations, such as performance variations across languages due to differences in training data volume, potential instability with very long texts or highly expressive inputs, and generation-to-generation variance in audio design and style control outputs. Note that these performance evaluations and limitations are published information for the base model “openbmb/VoxCPM2", and this “ggml-org/VoxCPM2-GGUF" is a GGUF model generated via an automated conversion tool (ggml-org/convert).

Strengths and Use Cases

The base model “openbmb/VoxCPM2" excels at flexible, high-quality speech synthesis and voice cloning across multiple languages. The main use cases and feature highlights are as follows:

  • Direct Multilingual Speech Synthesis: Supports 30 languages including Japanese, English, Chinese, Korean, French, German, and Spanish, as well as 9 Chinese dialects. Without requiring individual language tags, it automatically infers appropriate prosody and expressiveness from the input text to output speech.
  • Voice Design via Text Descriptions: Without preparing a reference audio file, new voice timbres can be generated simply by describing characteristics such as gender, age group, vocal quality, emotion, and speaking rate in natural language within parentheses at the beginning of the text, such as “(A young woman, gentle and sweet voice)".
  • Controllable Voice Cloning: A speaker’s vocal quality can be replicated by providing a short reference audio clip. It is also possible to change speaking speed, tone, and emotional expression by adding style instructions via text descriptions while maintaining the speaker’s vocal quality. Transcripts are not required when specifying reference audio.
  • Ultimate Cloning: By providing accurate transcript text alongside the reference audio, users can perform high-precision cloning or speech continuation that faithfully reproduces subtle vocal nuances of the speaker.
  • Real-Time Streaming Synthesis: Supports streaming generation, which sequentially generates and outputs audio.
  • Fine-Tuning with Limited Data: Fine-tuning via LoRA or full SFT can be performed by preparing around 5 to 10 minutes of audio data.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 2.3B parameters (taken from the base model openbmb/VoxCPM2)

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) BF16 2.7GB 3.3GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as GGUF.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Recent Models in the Same Size Class

Models with up to 4B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
ggml-org/Soprano-1.1-80M-GGUF 80M 4GB apache-2.0 Soprano-1.1-80M-GGUF Lightweight English TTS Model: File List (2026-11-17)
autotrust/TTS-V3 1.9B 8GB component-specific-terms TTS-V3 Phone-Dialogue Specialized TTS Model: 8GB+ VRAM, File List (2026-11-17)
autotrust/TTS-V4 1.9B 8GB other TTS-V4 Chinese Text-to-Speech Model: 8GB+ VRAM, File List (2026-11-17)
phasefield-audio/Irodori-TTS-v4.1-Anime 766M 4GB mit Irodori-TTS-v4.1-Anime: Our Generated Audio, File List (2026-09-07)

Distributed Files

Weight files published in ggml-org/VoxCPM2-GGUF, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
mmproj-VoxCPM2-BF16.gguf 2.01GB
VoxCPM2-BF16.gguf 2.95GB
VoxCPM2-Q8_0.gguf 1.57GB

How to Get It

“ggml-org/VoxCPM2-GGUF" can be obtained in GGUF format from its Hugging Face repository. No special prior application or gated license agreement procedures are required to use it.

Using the llama.cpp toolset, direct execution from the terminal and operation as a server are both possible.

Example command for direct execution from the terminal

llama-tts -hf ggml-org/VoxCPM2-GGUF -p "(A calm male narrator)Hello world" --output out.wav

When performing voice cloning using a reference audio file, specify the audio file with the --tts-speaker-file option. Transcript data is not required in this case.

Example command to start in server mode

llama serve -hf ggml-org/VoxCPM2-GGUF

Once the server starts, speech synthesis functionality can be accessed via the /tts endpoint.

Note that this model is a GGUF model automatically converted using the ggml-org/convert script. For detailed usage documentation, please refer to the llama.cpp VoxCPM2 documentation.

Related Articles

What to Read Next

Sources