Nemotron-H 47B Base 8K
NVIDIA · 2025-04-12 · 46.8B parameters
Nemotron-H 47B Base 8K is NVIDIA's open-weight 47B hybrid Mamba-Transformer base model, released on April 12, 2025; the BF16 checkpoint contains exactly 46,791,554,816 parameters. NVIDIA produced it by pruning Nemotron-H 56B Base with its MiniPuzzle method and distilling it on 63 billion tokens in FP8, leaving 98 layers with five self-attention layers and an 8,192-token context, so that it can run in FP4 on a single 32 GiB GPU. Because its cumulative training exposure combines the parent's pretraining with the distillation run, no training-token total is stored; the weights are released under the NVIDIA Internal Scientific Research and Development Model License.
Benchmark scores
| Benchmark | Score |
|---|---|
| ARC-Challenge | 94.6 |
| Capability | 119.7 |
| GSM8K | 93.3 |
| HellaSwag (Unspecified Scoring Protocol) | 87.9 |
| HumanEval | 61.0 |
| HumanEval+ | 56.1 |
| MATH | 57.4 |
| MATH-500 | 57.9 |
| MBPP | 75.9 |
| MBPP+ | 65.6 |
| MMLU | 83.6 |
| MMLU Pro | 61.8 |
| OpenBookQA | 48.8 |
| PIQA | 83.9 |
| WinoGrande | 83.9 |