VoxCPM2
OpenBMB's 2B Apache 2.0 TTS: 30 languages, voice design from a text description, cloning and 48 kHz output on about 8 GB of VRAM.
In short. VoxCPM2 speaks 30 languages and nine Chinese dialects in 48 kHz studio-quality audio. You can create a new voice just by describing it, clone a voice from a short clip, or steer emotion and pace with instructions. It needs about 8 GB of GPU memory, and the Apache 2.0 license allows commercial use.
Summary
OpenBMB released VoxCPM2 in April 2026 as the latest major release of its VoxCPM text-to-speech line, which had a technical report in September 2025. It has about 2 billion parameters built on MiniCPM-4, was trained on more than 2 million hours of multilingual speech, and uses a tokenizer-free diffusion autoregressive design, so it does not depend on a discrete speech tokenizer. Its AudioVAE V2 takes 16 kHz audio in and writes 48 kHz audio out, and the language model runs at 6.25 Hz.
It supports 30 languages without a language tag, plus nine Chinese dialects including Sichuan, Cantonese, Wu and Minnan. Voice design creates a voice from a written description of gender, age, tone, emotion and pace. Controllable cloning copies a voice from a short clip with optional style guidance, and ultimate cloning continues a reference clip when you give its transcript. The card puts the real-time factor at about 0.30 on an RTX 4090, or 0.13 with Nano-vLLM, and the model needs about 8 GB of VRAM. LoRA and full fine-tuning work with as little as 5 to 10 minutes of audio.
The weights and code are Apache 2.0, free for commercial use. OpenBMB says voice design and style control can vary between runs, so generating one to three times is recommended, that quality depends on each language's training data, and that very long or highly expressive input can be unstable. It forbids impersonation, fraud and disinformation. Serving options include the voxcpm package, vLLM-Omni, and llama.cpp-omni with GGUF weights. The benchmark tables on GitHub are OpenBMB's own.
Good for
- Multilingual speech with 48 kHz output, including Chinese dialects
- Creating a new voice from a text description, without any reference audio
- Voice cloning and fine-tuning with a few minutes of audio, under Apache 2.0
- Local TTS on a GPU with about 8 GB of memory
Watch out for
- Voice design and style control vary between runs, so you may need to generate a few times
- Quality differs by language, and very long or highly expressive input can be unstable
- Impersonation, fraud and disinformation are strictly forbidden, and AI-generated audio should be labeled
- The benchmark results come from OpenBMB itself
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of VoxCPM2. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| VoxCPM2 BF16 | 2.29B | 4.96 GB | 8 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 8 GB of GPU memory.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- OmniVoice vs VoxCPM2: which open text-to-speech model should you use?
- Best open-source speech models for transcription and text-to-speech
- Best open-source text-to-speech models in 2026
- Open speech models: transcription is open, text-to-speech is mixed
Frequently asked questions
- What is VoxCPM2?
- VoxCPM2 speaks 30 languages and nine Chinese dialects in 48 kHz studio-quality audio. You can create a new voice just by describing it, clone a voice from a short clip, or steer emotion and pace with instructions. It needs about 8 GB of GPU memory, and the Apache 2.0 license allows commercial use.
- Can I use VoxCPM2 commercially?
- Yes. VoxCPM2 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is VoxCPM2 open source?
- VoxCPM2 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run VoxCPM2 on my own computer?
- Yes. The smallest variant, VoxCPM2, needs about 8 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was VoxCPM2 released?
- VoxCPM2 was released by OpenBMB on April 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- VoxCPM repository and news (GitHub) github.com
- VoxCPM2 model card (Hugging Face) huggingface.co
- VoxCPM documentation voxcpm.readthedocs.io
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| OmniVoice Xiaomi | 613M | – | CC BY-NC (Non-commercial) | 3.75 GB Laptop or phone | |
| Qwen3-TTS Alibaba Qwen | 910M – 1.92B | – | Apache 2.0 (Permissive) | 3.68 GB Laptop or phone | |
| MOSS-TTS OpenMOSS | 4.55B – 8.49B | – | Apache 2.0 (Permissive) | 10 GB Laptop or phone | |
| LFM3-Audio-2B Liquid AI | 2B | 32K | LFM Open 1.0 (Conditional) | 4.6 GB Laptop or phone |