Atlas

Models

← All models

Nemotron 3 Nano 30B A3B

NVIDIA · 2025-12-15 · 31.6B parameters

Nemotron 3 Nano 30B A3B is an open-weight NVIDIA MoE language model, trained from scratch as a unified model for reasoning and non-reasoning tasks; the released BF16 checkpoint contains exactly 31,577,937,344 parameters. Its 52-layer hybrid architecture uses no positional embeddings and combines 23 Mamba-2, 23 MoE, and 6 GQA layers, with each MoE layer activating 6 of 128 routed experts and using shared-expert capacity. After a reported 25T-token main pretraining run plus a reported 121B-token long-context stage, it underwent SFT, RLVR, and RLHF; reasoning can be toggled, and it supports up to 1M tokens although the Hugging Face config defaults to 262,144 due to VRAM requirements.

Benchmark scores

BenchmarkScore
AA-LCR33.7
AA-Omniscience Index-51.6
AIME 202591.0
Artificial Analysis Agentic Index2.0
Artificial Analysis Coding Index14.4
Artificial Analysis Intelligence Index14.2
Artificial Analysis Omniscience Accuracy17.1
Artificial Analysis Omniscience Hallucination Rate82.9
Artificial Analysis Openness Index83.3
BLXBench68.0
BLXBench Coding category73.6
BLXBench Cost category89.3
BLXBench Debugging category67.9
BLXBench Hallucination category71.4
BLXBench pass rate40.7
BLXBench Reasoning category67.5
BLXBench Refactoring category59.5
BLXBench Security category57.2
BLXBench Speed category94.4
BLXBench UI category33.1
BullshitBench v2 Leaderboard28.0
BuseyBench SVG2.3
Capability133.5
CritPt0.9
FormationEval77.4
GDPval-AA v2492
GDPval-AA v2 (normalized)0.0
Global MMLU Lite61.2
GPQA Diamond75.7
Humanity's Last Exam (Text-Only)10.2
Loading Atlas data…