BREAKING
NVIDIA Releases 'TwoTower' Diffusion LM
0x
throughput
0%
quality kept
0T
training tokens
Two Towers Split the Work
Context Towerfrozen
Processes clean tokens causally
Reuses existing backbone
Denoiser Towertrained
Refines noisy blocks
Linked by cross-attention
Benchmarks vs AR Baseline
MMLU78.24
HumanEval75.58
GSM8K90.14
One Checkpoint, Three Modes
1Mask Diffusion
2Mock-AR
3Standard AR
Open Weights on Hugging Face
AI NEWS BLITZ
NVIDIA has unveiled TwoTower, a diffusion language model built for faster generation.