Atlas

Models

← All models

Nemotron 3 Nano Omni 30B A3B

NVIDIA · 2026-04-28 · 33B parameters

Nemotron 3 Nano Omni 30B A3B is NVIDIA's open-weight omni-modal reasoning model, combining a Nemotron 3 Nano 30B-A3B hybrid Mamba-2/Transformer MoE language backbone with a C-RADIOv4-H vision encoder, a Parakeet audio encoder, and multimodal projectors. It accepts text, image, video, and audio and produces text, supports up to 262,144 tokens, and allows reasoning to be toggled or constrained by an inference-time budget. The model uses Conv3D temporal compression for video and optionally supports runtime Efficient Video Sampling; this exact BF16 checkpoint contains 33,015,632,214 parameters and is governed by the NVIDIA Open Model Agreement.

Benchmark scores

BenchmarkScore
AA-LCR35.7
AA-Omniscience Index-56.0
Artificial Analysis Coding Index13.8
Artificial Analysis Intelligence Index14.9
Artificial Analysis Omniscience Accuracy14.8
Artificial Analysis Omniscience Hallucination Rate83.1
Artificial Analysis Openness Index55.6
Audio MultiChallenge (Text Output)23.2
BLXBench57.9
BLXBench Coding category62.8
BLXBench Cost category85.3
BLXBench Debugging category61.9
BLXBench Hallucination category60.8
BLXBench pass rate36.4
BLXBench Reasoning category51.9
BLXBench Refactoring category50.8
BLXBench Security category44.9
BLXBench Speed category87.2
BLXBench UI category17.6
BuseyBench SVG2.2
Capability131.4
CharXiv Reasoning63.6
CritPt0.0
GDP-PDF2.0
GDPval-AA v2465
GDPval-AA v2 (normalized)0.0
GPQA Diamond46.9
Humanity's Last Exam (Text-Only)5.3
IFBench (Artificial Analysis)63.2
MMMU Pro53.2
Loading Atlas data…