Full-duplex voice models you can run yourself
A full-duplex model listens and speaks at the same time, so you can interrupt it. NVIDIA’s NemotronLabs VoiceChat 11B is the first open one that calls tools mid-conversation, but it lists data-centre GPUs. MiniCPM-o 4.5 is the one that runs at home: OpenBMB asks for a 12 GB NVIDIA GPU or an M4 Max with 24 GB.
Written by Jakub Rusinowski · Last updated
The models
The licence column is the licence on the weights, read from each model card — not the badge on the code repository. Memory is what the maker or a named source publishes, not a measurement and not a fit verdict.
3 of 3 models
| Model | Hardware tier | Licence (weights) | Clones a voice | CPU only | Languages | Runtimes | Published memory |
|---|---|---|---|---|---|---|---|
| VoiceChat 11B NVIDIA · 11B | Server GPU (rent) | C OpenMDW-1.1 Commercial use allowed | — | No | — | vLLM NeMo Speech | Not published |
| MiniCPM-o 4.5 OpenBMB / ModelBest · 9B | 16–24 GB GPU / 32 GB+ Mac 8 GB GPU / 16 GB Mac | C Apache-2.0 Commercial use allowed | — | Yes | — | llama.cpp-omni Ollama (per card) vLLM | ~12 GB (published by OpenBMB / ModelBest (model card)) Full-duplex through llama.cpp-omni: an NVIDIA GPU with at least 12 GB, or an Apple M4 Max with at least 24 GB (card). llama.cpp-omni lists the full omni stack at Q4_K_M as ~8 GB model / ~9 GB peak on NVIDIA. |
| Moshi Kyutai | 16–24 GB GPU / 32 GB+ Mac | C CC-BY-4.0 — Attribution required; the card forbids impersonation. Commercial use allowed | No | No | — | official demo server | Not published |
No model matches these filters.
How to run it
MiniCPM-o 4.5 through llama.cpp-omni
OpenBMB publishes a C++ implementation with GGUF weights and an official Docker image for full-duplex use on a Mac; follow the guidance linked from the card. MiniCPM-o 4.5 card · llama.cpp-omni
NemotronLabs VoiceChat 11B through NeMo and vLLM
git clone https://github.com/NVIDIA-NeMo/Speech.gitFrom the model card, which lists A100, H100, H200, B100, B200 and RTX-6000 GPUs on Linux. VoiceChat 11B card
Announced, not downloadable yet
Dessa (Block): Announced, not downloadable: waitlist only, "launch links are coming soon". Trained from scratch on speech (no text, no pretrained models). Single source (Block). Listed so readers know it exists; excluded from the table filters. Block ↗
Sources: model cards on Hugging Face, the Open ASR Leaderboard, the Artificial Analysis Speech Arena and the Open TTS Leaderboard. Checked 25 October 2026.
Frequently asked questions
What is a full-duplex voice model?
A model that listens and speaks at the same time, so you can interrupt it mid-sentence and it can back-channel, instead of waiting for you to stop talking before it starts.
Can I run a full-duplex model on my own computer?
MiniCPM-o 4.5 runs through llama.cpp-omni on a 12 GB NVIDIA GPU or an Apple M4 Max with 24 GB, per OpenBMB’s model card. NemotronLabs VoiceChat 11B lists A100, H100, H200, B100, B200 and RTX-6000 class GPUs, so for most people that means renting one.
Is full duplex better than a speech-to-text, LLM and TTS chain?
It feels more natural and handles interruptions, but the chain lets you swap each piece and use any local LLM in the middle. Most local setups still use the chain.
What about Block’s Dessa?
Block has announced Dessa, a full-duplex speech-to-speech model trained from scratch on speech, and says it runs on a MacBook Pro M4 Max or better. As of 25 October 2026 there is only a waitlist: no weights and no licence, so it is not in the table.