Diffuplex
GitHubRelease pending
internal review.
Hugging FaceRelease pending
internal review.
arXiv Email

Full-duplex Spoken Language Modeling
with Discrete Diffusion

1,872Decode tokens / sGB300 · average throughput
13.0Tokens per forward passParallel decoding
86.9/84.1VoiceBench scoreThinking on/off
0.396 sFirst-audio latencyFull duplex interaction

HOW DIFFUPLEX WORKS

One set of parameters. Two modes.

AR=when to say

Listen continuously. Decide when to speak or stop.

Diffusion=what to say

Generate in parallel. Speak while refining the rest.

Listen → Trigger → Commit a prefix → Speak & refine

INTELLIGENCE MEETS FLUIDITY

Intelligence, at diffusion speed.

01 / THINKING ON

Competitive intelligence.
Parallel decoding.

≈3.9×

the decode throughput of the
fastest baseline shown

1,872 tokens/s

Decode throughput · GB300

86.9

VoiceBench intelligence score

ModelDecode throughput (tokens/s) →VoiceBench
Diffuplex
1,872
86.9
Phi-4MMFP8
481.7
64.3
N3-OmniFP8
393.6
89.4
Qwen2-AFP8
344.2
55.8
BR-VoiceBF16
327.3
90.46
Q3-Omni-TFP8
300.5
88.8
KimiBF16
293.3
76.9
MiniCPM-oBF16
289.2
71.2
Baichuan-1.5BF16
281.3
71.3
LFG-1BF16
254.2
79.6
GLM4-VBF16
214.0
56.5
LFG-3BF16
74.0
89.9
05001,0001,5002,000
GB300 · Fastest supplied configuration per baseline · Precision as indicatedVoiceBench: higher is better

02 / THINKING OFF · FULL-DUPLEX

Fast to respond.
Strong spoken intelligence.

0.396 s

First-audio latency · Fisher policy

82.3

VoiceBench intelligence score

+22.3 points over FLAIR, the highest-scoring full-duplex baseline in this comparison.

VoiceBench versus first-audio latency, zero-based score axis. Diffuplex: 82.3 at 0.396 s; FLAIR: 60.0 at 0.33; Nemotron-VC: 58.1 at 0.448; Freeze-Omni: 55.2 at 0.953; Lychee-FD: 43.5 at 0.84; SALM-Duplex: 39.0 at 0.55; PersonaPlex: 30.6 at 0.07; Moshi: 29.5 at 0.265.
Higher is smarter. Further left is faster.
From response modeling to full duplex.

VoiceBench: 86.9 with thinking, 84.1 without thinking at the end of Stage 2, and 82.3 after full-duplex adaptation. The Fisher policy improves response timing while retaining the same full-duplex intelligence score.

VoiceBench measures spoken language intelligence.View the original figure ↗

ACROSS THREE GPU PLATFORMS

The speed carries across hardware.

Decode throughput from the supplied measurements.

Decode tokens / s

Fastest supplied configuration per baseline; each GPU uses its own scale.Appendix figure ↗