Nemotron Nano 12B V2 VL BF16
NVIDIA · 2025-10-28 · 13.2B parameters
Nemotron Nano 12B V2 VL BF16 is the BF16 reference checkpoint of NVIDIA's Nemotron Nano V2 VL vision-language model, released on October 28, 2025; NVIDIA also ships FP8 (post-training-quantized) and NVFP4 (quantization-aware-distilled) variants. It combines the Nemotron Nano V2 12B hybrid Mamba-2/attention language backbone with a c-RADIOv2-H vision encoder, accepts text, multiple images and video, switches reasoning on or off, and supports a 131,072-token input-plus-output context. The checkpoint contains 13,181,860,358 parameters including the vision encoder and projector.
Benchmark scores
| Benchmark | Score |
|---|---|
| AIME 2025 | 71.3 |
| BLINK | 57.6 |
| Capability | 127.2 |
| ChartQA (test) | 89.8 |
| CharXiv Reasoning | 41.7 |
| CharXiv-D | 77.2 |
| DocVQA | 94.7 |
| GPQA Diamond | 64.1 |
| IFEval | 78.2 |
| IFEval (instruction-level strict accuracy) | 84.7 |
| LiveCodeBench - July 2024 to January 2025 | 69.4 |
| LongVideoBench - release-reported accuracy | 63.6 |
| MATH-500 | 96.9 |
| MATH-Vision | 53.6 |
| MathVista (mini) | 76.9 |
| MMBench v1.1 (English, split unstated) | 83.0 |
| MMLongBench-Doc | 32.1 |
| MMMU | 68.0 |
| MMStar | 71.7 |
| OCRBench | 856 |
| RealWorldQA - release-reported visual perception | 76.2 |
| ScreenSpot-Pro | 5.5 |
| ScreenSpot-v2 | 42.8 |
| TextVQA (validation) | 85.4 |
| Video-MME | 66.0 |