0

NVIDIA

GPU and AI-systems company supplying ~90% of the silicon used to train and serve frontier models; also publishes its own foundation models and eval suites.

Visit
Type
infra
HQ
Santa Clara, CA, USA
Founded
1993
Funding
Public (NASDAQ: NVDA); largest company in the world by market cap as of 2025
Website
nvidia.com
GitHub
github.com/NVIDIA
Twitter
twitter.com/nvidia

Cite

Notes

Only stored in your browser.

Evals
1
Tools
3
Models
24
Papers
2
Boards
0
People
16

Models

24

Ranked by Artificial Analysis Intelligence Index

#ModelReleasedIntelligence
1Nemotron 3 Super11 Mar 202636
2Nemotron Cascade 2 30B A3B19 Mar 202628.4
3Nemotron 3 Nano Omni 30B A3B Reasoning29 Apr 202621.4
4Llama 3.1 Nemotron Ultra 253B v1 (Reasoning)7 Apr 202515
5NVIDIA Nemotron 3 Nano 4B16 Mar 202614.7
6Llama Nemotron Super 49B v1.525 Jul 202514.6
7Llama 3.1 Nemotron Nano 4B v1.1 (Reasoning)20 May 202514.4
8Llama 3.3 Nemotron Super 49B v118 Mar 202514.3
9Llama 3.1 Nemotron Instruct 70B15 Oct 202413.4
10Nemotron 3 Nano 30B A3B15 Dec 202513.2
11Nemotron Nano 9B V218 Aug 202513.2
12Nemotron Nano 12B 2 VL28 Oct 202510.1
13NVIDIA-Nemotron-3-Super-120B-A12B-NVFP410 Mar 2026-
14Nvidia Llama 3.3 Nemotron Super 49B--
15Llama 3.3 Nemotron 49B Super--
16Nemotron 4 340B--
17Nvidia Nemotron 3 Nano 30B A3B Bf16--
18Llama 3.1 Nemotron 51B Instruct--
19Llama 3.1 Nemotron Ultra 253B--
20Nemotron 4 340B Instruct--
21Nvila Internal 15B--
22Nemotron 3 Super 120B A12B--
23Llama 3.3 70B Instruct FP4--
24Llama2 70B Steerlm Chat--
24 / 24 models

Evals

1

ComputeEval: CUDA Code Generation Benchmark

Evaluates LLM capability to generate correct CUDA code for kernel implementation, memory management, and parallel algorithm optimization tasks.

ActiveCoding

Tools

3

NeMo Gym Sokoban

NVIDIA's reference single-box Sokoban environment in NeMo Gym - a clean Gymnasium-style puzzle env used as the "hello world" for the NeMo RL training stack.

RL EnvPlanning

NVIDIA NeMo Gym

NVIDIA's RL-on-LLM framework inside NeMo - a library of agentic environments (code, math, search, games, robotics) that drop into the NeMo-RL trainer for large-scale, GPU-efficient post-training.

FrameworkCode GenerationMathPlanningAgentic

HelpSteer2

NVIDIA's permissively-licensed human-annotated preference dataset with 5-axis Likert ratings - engineered to train high-quality reward models.

PreferenceHallucinationSafetyInstruction Following

Papers

2

NeMo-Gym: NVIDIA's Framework for LLM Reinforcement Learning Environments

blog

2025

HelpSteer2: Open-source Dataset for Training Top-Performing Reward Models

NeurIPS

2024