Nemotron 3.5 Lightning 30B A3B Base
NVIDIA · 2026-08-11 · 32.9B parameters
Nemotron 3.5 Lightning 30B A3B Base is NVIDIA's pretrained (not post-trained) hybrid Mamba-2, attention and mixture-of-experts base model with multi-token prediction layers, the base of Nemotron 3.5 Lightning 30B A3B; it was released on Hugging Face on August 11, 2026. NVIDIA reports about 30B total and 3B active parameters, an NVFP4 pretraining recipe over more than 20T tokens, and up to 1M tokens of context. The checkpoint's safetensors metadata lists 32,913,266,240 parameters (BF16 plus F32 tensors).
Benchmark scores
| Benchmark | Score |
|---|---|
| Capability | 119.7 |
| GPQA Diamond | 33.3 |
| MATH-500 | 79.5 |
| MGSM Chinese | 78.4 |
| MGSM French | 82.4 |
| MGSM German | 84.8 |
| MGSM Japanese | 69.6 |
| MGSM Russian | 87.6 |
| MGSM Spanish | 88.0 |
| RULER - 1M | 69.6 |
| RULER - 256K | 76.9 |
| SuperGPQA | 33.2 |