Nemotron Nano 12B v2

NVIDIA Nemotron Nano 12B v2 is an open-weight reasoning and chat model released August 29, 2025, built on the Nemotron-Hybrid Mamba2-Transformer architecture: 12.31B parameters with just six attention layers, a 128K context length, and support for English plus German, French, Italian, Spanish, Japanese, and other languages. It supports runtime thinking-budget control, letting the user specify how many tokens the model may spend reasoning during inference.

On NVIDIA's reported benchmarks with reasoning on, it scores 76.25% on AIME25, 97.75% on MATH500, and 64.48% on GPQA, with evaluations reproducible through NeMo-Skills and a published tutorial. The weights are governed by the NVIDIA Open Model License Agreement and are ready for commercial use. A vision-language variant (Nemotron Nano 12B v2 VL) and FP4/FP8 checkpoints followed.