Atlas

Models

← All models

Llama 3.3 Nemotron Super 49B V1

NVIDIA · 2025-03-18 · 49.9B parameters

Llama 3.3 Nemotron Super 49B V1 is NVIDIA's open-weight reasoning model derived from Meta's Llama 3.3 70B Instruct and released on March 18, 2025; the released BF16 checkpoint contains exactly 49,867,145,216 parameters. Its 80-block decoder uses NAS-selected variable FFN widths, skips attention in 31 blocks, and uses grouped-query attention in the other 49. It was post-trained for switchable reasoning and non-reasoning chat, RAG, and tool calling, and supports a 131,072-token context window.

Benchmark scores

BenchmarkScore
AA-LCR17.0
AA-Omniscience Index-49.7
AIME 2024 (avg@8)61.3
AIME 2024-202553.5
AIME 202554.7
AIME 2025 (avg@8)45.8
Artificial Analysis Intelligence Index12.2
Artificial Analysis Omniscience Accuracy16.9
Artificial Analysis Omniscience Hallucination Rate73.0
Artificial Analysis Openness Index52.8
Capability130.5
CritPt0.0
GPQA Diamond64.6
GPQA Diamond (Vals protocol average)62.4
Humanity's Last Exam (Text-Only)6.5
IFBench (Artificial Analysis)39.5
LiveCodeBench (Artificial Analysis)28.0
LiveCodeBench (Vals AI public dataset)58.4
LiveCodeBench (Vals Public) - Easy95.0
LiveCodeBench (Vals Public) - Hard20.3
LiveCodeBench (Vals Public) - Medium59.8
MATH-50091.4
MATH-500 (Artificial Analysis source)95.9
MMLU Pro78.5
MMLU Pro - Biology86.1
MMLU Pro - Business82.6
MMLU Pro - Chemistry70.8
MMLU Pro - Computer Science76.3
MMLU Pro - Economics80.9
MMLU Pro - Engineering53.9
Loading Atlas data…