← All models

Nemotron Nano 9B V2 Base

NVIDIA · 2025-08-18 · 8.9B parameters

Nemotron Nano 9B V2 Base is NVIDIA's 9B hybrid Mamba-Transformer base (completion) model, produced by Minitron pruning (depth, FFN width and embedding channels) and knowledge distillation from Nemotron Nano 12B V2 Base, with a final distillation stage at 524,288-token sequences; it supports a 128K-token context. The BF16 checkpoint contains exactly 8,888,227,328 parameters in 56 layers with four GQA attention layers and no positional embeddings. Its distillation token counts are stage-specific, so no cumulative training-token figure is stored.

Benchmark scores