← All models

Nemotron Nano 12B V2

NVIDIA · 2025-08-29 · 12.3B parameters

Nemotron Nano 12B V2 is NVIDIA's aligned 12B hybrid Mamba-Transformer reasoning model, the post-trained Nemotron-Nano-12B-v2-Base checkpoint (SFT, GRPO, DPO, RLHF and checkpoint merging) that NVIDIA later pruned and distilled into Nemotron Nano 9B V2. 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, supports a 128K-token context, and switches reasoning on or off by system prompt with optional thinking-budget control. Its base was pretrained on about 20 trillion tokens with an FP8 recipe, so the checkpoint's cumulative training exposure is not stored as an exact count.

Benchmark scores