nvidia/personaplex-7b-v1

Audio-to-Audio • 8B • Updated Mar 2 • 175k • • 2.78k

Note Full-duplex speech conversation with voice and role prompts, built on Moshi. English. Gated; NVIDIA Open Model License.

kyutai/moshiko-pytorch-bf16

8B • Updated Sep 18, 2024 • 176k • 258

Note Original Moshi-family full-duplex speech conversation baseline; Moshiko voice, PyTorch BF16 checkpoint. English. CC-BY-4.0.

kyutai/personaplex-rl-seamless

Audio-to-Audio • 8B • Updated Jun 11 • 4.5k • 38

Note PersonaPlex post-trained on Seamless Interaction to improve pause handling, turn-taking, backchanneling, and interruption behavior.

openbmb/MiniCPM-o-4_5

Any-to-Any • 9B • Updated Aug 18 • 735k • 1.5k

Note Multimodal full-duplex model: concurrent audio/video input and text/speech output. Use its duplex inference mode; also supports separate half-duplex modes.

HIT-TMG/Lychee-FD

13B • Updated Aug 20 • 330 • 5

Note Research full-duplex speech checkpoint for Chinese/English. Demo requires the separate Token2Wav vocoder from Step-Audio-2-mini. Apache-2.0 metadata.

BayLing-Models/BayLing-Duplex

516k • Updated Jun 15 • 189 • 7

Note Research model with native listening/speaking and learned interruption/turn timing. Requires separate GLM-4-Voice tokenizer and decoder; see LICENSE/NOTICE for terms.

MuyeHuang/DuplexOmni

35B • Updated Jul 26 • 991 • 3

Note Research multimodal full-duplex Thinker/Talker model. Card notes unexpected silence/speech-quality issues and recommends at least 8 H20 GPUs for low latency. Apache-2.0.

mindlogicinc/context-spanning-7b

Audio-to-Audio • Updated 5 days ago

Note Recent PersonaPlex-derived full-duplex research model. Published runtime uses external router-LLM and ASR backends. NVIDIA Open Model License; not a standalone deployment.

nu-dialogue/j-moshi

8B • Updated Jun 4, 2025 • 59 • 15

Note Japanese full-duplex Moshi adaptation with overlapping speech and backchannels. Official research checkpoint; j-moshi-ext is the companion variant trained with additional synthetic dialogue.

VITA-MLLM/Freeze-Omni

Updated Nov 26, 2024 • 21

Note Duplex speech system with chunk-level interruption/state prediction and streaming input/output. Requires separate Qwen2-7B-Instruct weights; server uses VAD and cache scheduling.

gpt-omni/mini-omni2

Any-to-Any • Updated Oct 24, 2024 • 138 • 289

Note Early multimodal duplex research baseline with end-to-end streaming speech output. Paper demonstrates command-based interruption; this is narrower than unrestricted conversational overlap handling.

tsinghua-ee/ELLSA

18B • Updated Apr 19 • 50 • 5

Note End-to-end streaming full-duplex vision, speech, text and action model. Includes speaking-while-acting and action barge-in; research setup requires multiple companion checkpoints.

sbintuitions/DuplexCascade

Text Generation • Updated Aug 31 • 4

Note Cascaded full-duplex system checkpoint: Qwen2-7B-Instruct fine-tuned for VAD-free micro-turn dialogue control. Requires external ASR and TTS; this repository is the LLM component.