Latest29 Sep 2026Embedding/other
NVIDIA Kumo Tabular
An open foundation model for prediction on tables. Given a table with labelled rows and the rows to predict, it returns class probabilities or numeric predictions in a single forward pass, with no training, tuning or feature engineering. It extends NVIDIA's open models beyond language and vision into structured business data.
- Tabular classification and regression by in-context learning, no training
- Three sizes, 28M to 215M parameters (NVIDIA)
- Pretrained only on artificial data (NVIDIA)
- First overall on TabArena, ELO 1950 (NVIDIA-reported)
AccessOpen weights on Hugging Face (nvidia/Kumo-Tabular) under the OpenMDW-1.1 licence, which NVIDIA says allows commercial use; runs through NVIDIA's open-source structured-data library
11 Aug 2026Open-weights LLM
NVIDIA Nemotron 3.5 Lightning
Nemotron 3.5 Lightning is a 30-billion-parameter open model from NVIDIA that uses about 3 billion parameters per token, built for fast, high-volume steps inside AI agents. It mixes Mamba-2, mixture-of-experts and attention layers and was trained for speculative decoding. It runs on RTX PCs, DGX Spark and Jetson as well as data centres.
- 30B total / 3B active, hybrid Mamba-2 MoE
- Up to 1M-token context (256K on one H100)
- SWE-bench Verified 51.56%, GPQA Diamond 75.44 (model card)
- BF16 and NVFP4 checkpoints plus draft models
AccessOpen weights, data and recipes under OpenMDW-1.1 on Hugging Face; build.nvidia.com, OpenRouter, ModelScope; Amazon SageMaker, Google Cloud, Microsoft Foundry, OCI
4 Jun 2026Open-weights LLM
NVIDIA Nemotron 3 Ultra
Nemotron 3 Ultra is NVIDIA's largest open model, a 550-billion-parameter mixture-of-experts with 55 billion active parameters, built for long-running agents that reason, plan and use tools. NVIDIA released the weights, training data and recipes under the permissive OpenMDW licence and names it the flagship of the Nemotron 3 family.
- 550B total / 55B active, Mamba-2 + MoE hybrid
- Up to 1M-token context; 10 languages
- SWE-Bench Verified 70.7%, MMLU-Pro 86.8% (model card)
- Reasoning can be switched on or off
AccessOpen weights, data and recipes under OpenMDW-1.1 on Hugging Face; build.nvidia.com, Perplexity, OpenRouter and cloud providers
1 Jun 2026World model
NVIDIA Cosmos 3
Cosmos 3 is NVIDIA's open world foundation model family for physical AI, combining vision reasoning, world generation and action prediction in one mixture-of-transformers model. It is aimed at robots, self-driving cars and vision agents, including generating physics-accurate synthetic training data. Super and Nano versions launched at Computex; an Edge version is coming.
- Understands and generates text, images, video, sound and actions
- Physics-aware world simulation for synthetic data
- Variants: Cosmos 3 Super and Nano; Edge coming soon
AccessOpen models on Hugging Face, GitHub, build.nvidia.com and NVIDIA NIM; Cosmos 3 Super and Nano available, Edge coming
28 Apr 2026Open-weights LLM
NVIDIA Nemotron 3 Nano Omni
Nemotron 3 Nano Omni is NVIDIA's open multimodal model that understands video, audio, images and text in one 31-billion-parameter network with about 3 billion active parameters. It is aimed at enterprise Q&A, summaries, transcription and document work, and fits on a single H100. NVIDIA first announced it as Nemotron 3 Omni at GTC 2026.
- Inputs: video, audio, image and text; outputs text
- 256K-token context window
- OSWorld computer use 47.4% (model card)
- BF16, FP8 and NVFP4 checkpoints
AccessOpen weights on Hugging Face (NVIDIA Open Model Agreement, commercial use permitted), build.nvidia.com and NGC
17 Apr 2026Embedding/other
Isaac GR00T N1.7
Isaac GR00T N1.7 is NVIDIA's 3-billion-parameter vision-language-action model for humanoid robots, separating high-level reasoning from motor control. It was pre-trained on about 20,000 hours of first-person human video to improve dexterous hand control. NVIDIA describes it as the first GR00T version with a commercial licence.
- 3B-parameter vision-language-action model for humanoids
- Finger-level manipulation with 22-degree-of-freedom hands
- Works across Unitree G1, AGIBot Genie 1 and YAM
- Breaks multi-step tasks into sub-steps
AccessOpen weights on Hugging Face (nvidia/GR00T-N1.7-3B); code on GitHub (Apache 2.0); early access at launch, later marked general availability
16 Mar 2026Small/fast LLM
NVIDIA Nemotron 3 Nano 4B
Nemotron 3 Nano 4B is a 4-billion-parameter open model from NVIDIA for AI agents on edge devices such as Jetson Thor, GeForce RTX PCs and DGX Spark. Its hybrid design is mostly Mamba-2 and MLP layers with only four attention layers, and it accepts inputs of up to about 262,000 tokens.
- About 3.97B parameters, built for edge devices
- Up to 262K-token context
- MATH500 95.4 with reasoning off (model card)
- BF16, FP8 and GGUF releases
AccessOpen weights on Hugging Face (NVIDIA Nemotron Open Model License); BF16, FP8 and GGUF
11 Mar 2026Open-weights LLM
NVIDIA Nemotron 3 Super
Nemotron 3 Super is the mid-sized model in NVIDIA's open Nemotron 3 family, with 120 billion total and 12 billion active parameters in a hybrid Mamba-Transformer mixture-of-experts design. It is aimed at agent reasoning and coding with efficient inference, and accepts up to 1 million tokens of context.
- 120B total / 12B active, latent mixture-of-experts
- Up to 1M-token context
- AIME25 90.21% without tools (model card)
- SWE-Bench 60.47% via OpenHands (model card)
AccessOpen weights on Hugging Face (NVIDIA Nemotron Open Model License); BF16, FP8 and NVFP4
15 Dec 2025Open-weights LLM
NVIDIA Nemotron 3 Nano
Nemotron 3 Nano was the first model in NVIDIA's Nemotron 3 family: a 30-billion-parameter open model with about 3.5 billion active parameters, mixing Mamba-2, mixture-of-experts and attention layers. NVIDIA says it delivers four times the throughput of Nemotron 2 Nano, and it accepts up to 1 million tokens of context.
- 30B total / ~3.5B active hybrid MoE
- Up to 1M-token context (256K default)
- AIME25 99.2% with tools; SWE-Bench 38.8% (model card)
- 4x throughput of Nemotron 2 Nano (NVIDIA)
AccessOpen weights on Hugging Face (NVIDIA Nemotron Open Model License); BF16, FP8 and NVFP4