NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4

Home/Models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4

NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4

nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4

kind: nvfp4
quant: nvfp4 (4-bit)
family: nemotron
modality: text-generation
root model: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
provider: nvidia
license: other
language: en, fr, es, it, de, ja, zh
context: 1M
library: transformers
vRAM (GB): 812.7
size on disk: —
downloads: 3,017,149
likes: 402
created: —
last modified: 2026-05-01T16:54:19+00:00 (2026)
huggingface: https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4
tags: transformerssafetensorsnemotron_htext-generationnvidiapytorchnemotron-3latent-moemtpconversationalcustom_codeenfresitdejazhdataset:nvidia/nemotron-post-training-v3dataset:nvidia/nemotron-pre-training-datasetsarxiv:2512.20848arxiv:2512.20856license:otherendpoints_compatible

Benchmark scores (13)

benchmark	category	score	context	conf	evidence
RULER	long-context	96.79	RULER-500 @ 128k (500 samples per task)	medium	`\| RULER-500 @ 128k (500 samples per task) \| 96.79 \| 96.85 \| 95.99 \|`
RULER	long-context	96.6	RULER-500 @ 256k (500 samples per task)	medium	`\| RULER-500 @ 256k (500 samples per task) \| 96.60 \| 96.33 \| 96.52 \|`
RULER	long-context	96.09	RULER-500 @ 512k (500 samples per task)	medium	`\| RULER-500 @ 512k (500 samples per task) \| 96.09 \| 95.66 \| 96.23 \|`
MMLU-Pro	knowledge	83.73	MMLU-Pro	medium	`\| MMLU-Pro \| 83.73 \| 83.63 \| 83.33 \|`
LiveCodeBench	coding	81.19	LiveCodeBench (v5 2024-07↔2024-12)	medium	`\| LiveCodeBench (v5 2024-07↔2024-12) \| 81.19 \| 80.99 \| 80.56 \|`
MMLU-Pro	knowledge	79.35	MMLU-ProX (avg over languages)	medium	`\| MMLU-ProX (avg over languages) \| 79.35 \| 79.21 \| 79.37 \|`
GPQA	reasoning	79.23	GPQA (no tools)	medium	`\| GPQA (no tools) \| 79.23 \| 79.36 \| 79.42 \|`
LiveCodeBench	coding	78.69	LiveCodeBench (v6 2024-08↔2025-05)	medium	`\| LiveCodeBench (v6 2024-08↔2025-05) \| 78.69 \| 78.44 \| 78.44 \|`
Chatbot Arena	chat	73.88	Arena-Hard-V2 (Hard Prompt)	medium	`\| Arena-Hard-V2 (Hard Prompt) \| 73.88 \| 76.06 \| 76.00 \|`
IFBench	instruction-following	72.58	IFBench (prompt)	medium	`\| IFBench (prompt) \| 72.58 \| 72.32 \| 73.30 \|`
SciCode	coding	42.05	SciCode (subtask)	medium	`\| SciCode (subtask) \| 42.05 \| 41.38 \| 40.83 \|`
Terminal-Bench	agentic-coding	25.78	Terminal Bench (hard subset)	medium	`\| Terminal Bench (hard subset) \| 25.78 \| 26.04 \| 24.48 \|`
Humanity's Last Exam	knowledge	18.26	HLE (no tools)	medium	`\| HLE (no tools) \| 18.26 \| 17.42 \| 17.42 \|`

Discussions (34)

title	author	status	type
TemporalMesh Transformer: 29.4 PPL at 48% compute — beats Mamba, new open-source architecture	vigneshwar234	open	disc
Can't run on dgx spark with vllm	moranilt	open	disc
Model size is halved for NVFP4	YouNeedCryDear	open	disc
nemotron120b	willowoods	draft	disc
Add streaming reasoning extraction with content promotion	avskliar-nvidia	open	disc
Fix: rescue whitespace-only final_content (e.g. "\n"), not just None	avskliar-nvidia	merged	disc
Update Inference Configs	Chris-Alexiuk	merged	disc
tool call leaks	kristianpaul	open	disc
Update vLLM guidance.	llm-wizard	merged	disc
Update to Docker Based Spark Command	llm-wizard	merged	disc
Update README.md	llm-wizard	merged	disc
--reasoning-config breaks Nemotron v3 reasoning parser (content always null, thinking unbounded)	rhxsec	open	disc
"This will lead to incorrect tokenization" warning	DanTup	open	disc
Update Instruction Clarity	llm-wizard	merged	disc
Jetson Thor Official Container for vLLM 0.16 fails to load nemotron-3-super -- says mixed-precision quant config is unsupported in vLLM 0.16 container	mrjbj	open	disc
FP4 quantization for inference optimization	O96a	open	disc
Spark not using NVFP4?	D-Lynch	open	disc
Update Snippets	Chris-Alexiuk	merged	disc
VLLM + MTP + NVFP4 doesn't work	catplusplus	open	disc
Searching for a new Tool Parser	LucasMM14	open	disc
Run on DGX Spark	LimeemiL	open	disc
All this talk about NVFP4 - why is it dog slow?	josephbreda	open	disc
NVFP4 cannot be loaded in SGLang	mratsim	open	disc
Update the usage instructions for SGLang 1M context	shuwennn	merged	disc
Updated README to fix wget link for reasoning_parser.py	rohit-nvidia	merged	disc
vLLM MTP unusable on RTX 6000 Pro, as spec decoding consumes 20GB+ VRAM at start-up, causing OOM	lsmc	open	disc
Doesn't work with latest vllm, even tried to recompile vLLM and transformers from git	catplusplus	open	disc
RTX Pro 6000 support	justinjja	open	disc
CUDA Version -- Min requirement?	raymondlo84-nvidia	open	disc
Fix `wget` command in readme	DanTup	closed	disc
Update README.md	mmangkad	closed	disc
Update README.md	dmax123	merged	disc
Create README.md	dmax123	merged	disc
Update super_v3_reasoning_parser.py	dmax123	merged	disc