← All models

Nemotron 3 Ultra 550B A55B Base

NVIDIA · 2026-06-04 · 560.5B parameters

Nemotron 3 Ultra 550B A55B Base is NVIDIA's pretrained (not post-trained) LatentMoE hybrid Mamba-2/Transformer base model with multi-token prediction layers, the base of Nemotron 3 Ultra 550B A55B. The BF16 checkpoint contains exactly 560,524,578,816 parameters with 512 routed experts of which 22 are active per token; NVIDIA reports about 55B active parameters. The card says it was pretrained on around 20T tokens with an NVFP4 recipe and supports up to 1M tokens of context; the approximate token count is not stored as exact.

Benchmark scores