Atlas

Models

← All models

Nemotron 3 Nano 4B

NVIDIA · 2026-03-16 · 4B parameters

Nemotron 3 Nano 4B is NVIDIA's open-weight, text-only small language model compressed from NVIDIA-Nemotron-Nano-9B-v2 with the Nemotron Elastic framework and released as a BF16 checkpoint for both reasoning and non-reasoning tasks. The checkpoint contains exactly 3,973,556,832 parameters, and its 42-layer hybrid architecture comprises 21 Mamba-2, four grouped-query-attention, and 17 MLP layers. Reasoning is switchable through the chat template, and the released configuration supports a 262,144-token context window.

Benchmark scores

BenchmarkScore
AA-LCR16.7
AA-Omniscience Index-71.5
Artificial Analysis Agentic Index2.3
Artificial Analysis Coding Index8.0
Artificial Analysis Intelligence Index8.7
Artificial Analysis Omniscience Accuracy9.4
Artificial Analysis Omniscience Hallucination Rate89.4
Artificial Analysis Openness Index83.3
Capability127.6
CritPt0.0
GDPval-AA v2204
GDPval-AA v2 (normalized)0.0
GPQA Diamond51.3
Humanity's Last Exam (Text-Only)4.8
IFBench (Artificial Analysis)58.2
SciCode (Artificial Analysis)16.4
Terminal-Bench 2.13.7
Terminal-Bench Hard6.8
τ²-bench (Telecom)28.1
𝜏³-Banking7.0
Loading Atlas data…