NVIDIA releases a Nemotron model that fills token blocks in parallel

NVIDIA released Nemotron-TwoTower-30B-A3B-Base-BF16. It is a built from the Nano 30B-A3B backbone.

Most write by producing one token after another in order. This model combines a frozen context tower with a diffusion denoiser tower, then fills blocks of tokens in parallel through repeated passes.

In NVIDIA's default mask-diffusion setup, it kept 98.7% of the aggregate benchmark quality of the baseline. Its wall-clock generation reached 2.42 times that baseline.

Key points

  • Nemotron-TwoTower-30B-A3B-Base-BF16 is based on Nano 30B-A3B.
  • It does not only generate one token at a time; it fills token blocks in parallel.
  • NVIDIA reported 98.7% of the baseline's aggregate benchmark quality.
  • NVIDIA reported 2.42 times the wall-clock generation of that baseline.
  • For AI agents, the main promise is lower latency and possibly lower .
Read original