Atlas

Models

← All models

Nemotron Nano 9B V2

NVIDIA · 2025-08-18 · 8.9B parameters

Nemotron Nano 9B V2 is NVIDIA's 9B-class hybrid Mamba-Transformer language model, released in August 2025 and produced by pruning and distilling a 12B base model pretrained on approximately 20 trillion tokens. Its architecture replaces most self-attention layers with Mamba-2 layers to improve inference throughput on long reasoning traces, with reasoning behavior controllable via system prompt. Because the base model's training-token figure is approximate and does not describe this checkpoint's cumulative exposure, it is not stored as an exact count.

Benchmark scores

BenchmarkScore
AA-LCR22.7
AA-Omniscience Index-42.9
AIME 202569.7
Artificial Analysis Intelligence Index8.8
Artificial Analysis Omniscience Accuracy11.3
Artificial Analysis Omniscience Hallucination Rate61.0
Artificial Analysis Openness Index72.2
BullshitBench v2 Leaderboard5.0
BuseyBench SVG1.2
Capability129.9
CritPt0.0
Fiction.liveBench25.0
FormationEval79.6
GPQA Diamond57.0
Humanity's Last Exam (Text-Only)4.6
IFBench (Artificial Analysis)27.6
LiveCodeBench (Artificial Analysis)72.4
MMLU Pro74.2
SciCode (Artificial Analysis)22.0
SnakeBench Average Apples1.1
SnakeBench Best Apples15.0
SnakeBench Rating18.0
SnakeBench Total Apples51.0
SnakeBench Win Rate45.5
Terminal-Bench Hard1.5
τ²-bench (Telecom)23.4
Loading Atlas data…