← All models

Nemotron 3 Nano 30B A3B Base

NVIDIA · 2025-12-15 · 31.6B parameters

Nemotron 3 Nano 30B A3B Base is NVIDIA's pretrained (not post-trained) hybrid Mamba-2/Transformer mixture-of-experts base model, the starting point for Nemotron 3 Nano 30B A3B and Nemotron Cascade 2. The BF16 checkpoint contains exactly 31,577,937,344 parameters in 52 layers (23 Mamba-2, 23 MoE and 6 GQA attention layers) with 128 routed experts of which 6 are active per token, and no positional embeddings. NVIDIA documents support for up to 1M tokens of context although the Hugging Face configuration defaults to 256K; the card lists a 128K maximum input size.

Benchmark scores