Model
Model explorer

Llama-3.3-Nemotron-Super-49B v1 (Reasoning Off)

OPEN
NVIDIA · Llama Nemotron family · released Mar 18, 2025

Toggleable reasoning mode via system prompt; NAS-compressed to fit a single H100/H200 at high throughput.

ReasoningCodingVisionFunction callingTool useAgentic
1116.2
Elo · rank #188
Parameters
49.9B
Active params
49.9B (dense)
Context
128K tokens
Architecture
Dense Transformer, NAS-compressed from Llama-3.3-70B-Instruct (variable FFN + skip-attention blocks)
License
NVIDIA Open Model License + Llama 3.3 Community License Agreement
Languages
8+
API price (in/out)
$0.1 / $0.5
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
AIMEMath13.3%#145
best: GPT-5.2 · 100.0%
Arena-HardHuman preference88.3%#15
best: Qwen3-235B-A22B (Non-Thinking) · 96.1%
BFCLAgents73.7%#19
best: SmolLM3 3B (No Thinking) · 92.3%
GPQA DiamondReasoning50.0%#191
best: GPT-5.6 · 94.6%
Humanity's Last ExamReasoning4.0%#130
best: Claude Sonnet 5 · 57.4%
IFBenchReasoning39.0%#49
best: MiniMax M3 · 83.0%
IFEvalReasoning89.2%#26
best: Gemma 4 26B A4B · 98.5%
LiveCodeBenchCoding29.7%#150
best: DeepSeek-V4-Pro (Think Max) · 93.5%
MATH-500Math74.0%#81
best: GPT-5 · 99.4%
MBPPCoding84.9%#4
best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%
MT-BenchHuman preference9.17#6
best: Hunyuan-Large (A52B) · 9.4
Run it locally
VRAM @ Q4
30 GB
VRAM @ FP16
99.7 GB
Fits on (Q4)
M4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
~30GB at Q4 exceeds a single RTX 4090's 24GB; no verified 4090 figure found.
Quantizations
GGUF · MLX · FP8
Llama Nemotron family
Elo progression across releases
Llama-3.1-Nemotron-70B InstructOct 2024661.5Llama-3.3-Nemotron-Super-49B v1 (Reasoning Off)Mar 20251116.2Llama-3.3-Nemotron-Super-49B v1 (Reasoning On)Mar 20251378.0Llama-3.1-Nemotron-Ultra-253B v1 (Reasoning Off)Apr 20251120.4Llama-3.1-Nemotron-Ultra-253B v1 (Reasoning On)Apr 20251544.6
Resources
HFnvidia/Llama-3_3-Nemotron-Super-49B-v1DOCSdocs.api.nvidia.com/nim/reference/nvidia-llama-3_3-nemotron-super-49b-v1APIAvailable via NVIDIA API
API price $0.1/$0.5 · each benchmark row carries its own source badge (see methodology)