NVIDIA / FULL-DUPLEX SPEECH / 2026

PersonaPlex-7B-v1

TOTAL8.37Bparameters

ACTIVE8.37Bruntime path

DEPTH32repeated blocks

HIDDEN4,096main trunk

CONTEXTstreamstream / sequence

CHECKPOINT17.14GB2 shards · NVIDIA Open Model

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

30 seconds → 30 seconds · 17.03T arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT30SECONDS

TRANSFORMER STACK32 SPEECH LM LAYERSsimultaneous user-listen and assistant-speak streams

Multistream Residual

OUTPUT30SECONDS

wraps every repeated block

HELIUM TRANSFORMER · FULL-DUPLEX TURNTAKING模型不等待完整轮次结束，能够同时感知打断并继续生成。

Mimi CodebooksMimi codec 把双向 waveform 压为实时离散 token 流。

DEPTH32speech LM

HIDDEN4,096trunk

CODEBOOKS8Mimi

AUDIO24kHzmono

MODEduplexreal time

PARAMS8.37Bheader