Atlas

Models

← All models

Nemotron Nano 12B V2 VL

NVIDIA · 2025-10-28 · 13.2B parameters

Nemotron Nano 12B V2 VL FP8 is NVIDIA's open-weight, post-training-quantized E4M3 checkpoint of the Nemotron Nano V2 VL vision-language model, released on October 28, 2025. It combines the Nemotron Nano V2 12B hybrid Mamba-2/attention language backbone with a c-RADIOv2-H (RADIOv2.5-H) vision encoder and supports text, image, and video inputs, switchable reasoning-on/reasoning-off modes, and a released 131,072-token input-plus-output context; the final training stage used sequences up to 311,296 tokens. It targets document understanding, OCR, visual question answering, image summarization, and long-video comprehension.

Benchmark scores

BenchmarkScore
AA-LCR40.0
AA-Omniscience Index-64.4
AIME 202575.0
Artificial Analysis Intelligence Index9.0
Artificial Analysis Omniscience Accuracy13.8
Artificial Analysis Omniscience Hallucination Rate90.7
Artificial Analysis Openness Index72.2
BuseyBench SVG1.3
Capability129.5
CritPt0.0
FormationEval77.4
GPQA Diamond57.2
Humanity's Last Exam (Text-Only)5.3
IFBench (Artificial Analysis)31.9
LiveCodeBench (Artificial Analysis)69.4
MMLU Pro75.9
MMMU Pro52.9
SciCode (Artificial Analysis)26.2
Terminal-Bench Hard4.5
τ²-bench (Telecom)21.3
Loading Atlas data…