← All models

Nemotron Nano 12B V2 Base

NVIDIA · 2025-08-18 · 12.3B parameters

Nemotron Nano 12B V2 Base is NVIDIA's 12B hybrid Mamba-Transformer base (completion) model, pretrained from scratch on about 20 trillion tokens of curated and synthetic data with an FP8 recipe and a warmup-stable-decay schedule, followed by a long-context phase that extends it to a 128K-token context. The BF16 checkpoint contains exactly 12,310,001,152 parameters in 62 layers (28 Mamba-2, 28 MLP and 6 GQA attention layers) with no positional embeddings. It is the parent of the aligned Nemotron Nano 12B V2 and of the pruned Nemotron Nano 9B V2 Base.

Benchmark scores