NVFP4
NVIDIA's 4-bit microscaling format, using small value blocks with a floating-point scale; the choice between it and MXFP4 is baked into silicon and into checkpoint portability.
Current numbers
1st rack validated 2026-06-01CoreWeave-reported industry-first Vera Rubin NVL72 rack, validated 1 June 2026: ~3.6 EFLOPS sparse NVFP4, 216 TB/s NVLink 6
25T tokenslongest documented 4-bit pre-training run by tokens (Nemotron 3 Super, 120B total, ~12B active MoE, NVFP4, 25T total seen tokens)
~50 PFLOPS sparse (35 dense)Vera Rubin per-GPU NVFP4: 50 PFLOPS sparse (inference), 35 PFLOPS dense (training)