NVIDIA Isaac GR00T N1.7 3B
Release in the NVIDIA Isaac GR00T N family · version N1.7 (GR00T-N1.7-3B)
GR00T N1.7 is a 3B-parameter vision-language-action model from NVIDIA that maps camera images, language instructions, and robot proprioception to continuous robot actions. It replaces the earlier Eagle backbone with Cosmos-Reason2-2B and uses a flow-matching diffusion transformer as its action head. NVIDIA first released it in early access in April 2026; the repository now describes it as a general availability release.135
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- Repository: Isaac-GR00T repository (external site: github.com)
- License: NVIDIA Open Model License (external site: nvidia.com)
- Release notes: GR00T N1.7 announcement (Hugging Face blog) (external site: huggingface.co)
Availability and license
Overall availability
Weights are downloadable from Hugging Face without an access gate under the NVIDIA Open Model License; the card states the model is ready for commercial and non-commercial use.21
Availability is separate from permission: read the license before using or redistributing.
NVIDIA Open Model License Agreement (external site: nvidia.com)136
Apache License 2.0 (Isaac-GR00T code) (external site: raw.githubusercontent.com)43
The NVIDIA Open Model License allows commercial use, modification, and distribution of the model and derivatives, with NVIDIA claiming no ownership of outputs. It requires use consistent with NVIDIA's Trustworthy AI terms, a notice file with a specified NVIDIA attribution on redistribution, and indemnification of NVIDIA; rights terminate for bypassing safety guardrails or for bringing infringement litigation over the model. It is governed by U.S. and Delaware law. The README's overview also says N1.7 is licensable under Apache 2.0, but its license section and the model card assign Apache 2.0 to the code and the NVIDIA Open Model License to the weights.631
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
The weights are under a license that is not on the rubric's OSI-approved list. Read its terms before use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | Base weights and several post-trained benchmark checkpoints are published on Hugging Face without an access gate.21 |
| Inference codeIs code for running the model published? | Public | The Apache-2.0 Isaac-GR00T repository provides zero-shot and fine-tuned inference, a policy server, and ONNX/TensorRT export.34 |
| Training codeIs the code used to train the model published? | Unknown | The repository publishes fine-tuning code (launch_finetune.py) for new embodiments; pre-training code for the base model was not found.3 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The card describes the pre-training mixture at a high level (13 datasets, 21.6 million data points, human, robot, and simulated data), and NVIDIA's announcement names 20,854 hours of EgoScale human egocentric video. The full pre-training data is not documented as released.15 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The card points to the GR00T N1 white paper for the N1.x architecture, and the README summarizes the N1.7 changes (new backbone, relative end-effector action space, human video pre-training); a full N1.7 training configuration was not found.13 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Public | The README links per-benchmark guides with dataset download, fine-tuning, and evaluation commands for reproducing results on LIBERO, SimplerEnv, and SO100, plus open-loop and closed-loop evaluation tools.3 |
What it is useful for
Run and use notes
- The card lists PyTorch as the runtime and NVIDIA Ampere, Hopper, Lovelace, Blackwell, and Jetson hardware on Linux, and reports inference latency with PyTorch eager mode, torch.compile, and TensorRT on several GPUs, DGX Spark, and Jetson AGX Thor and Orin.1
Organization context
Provenance and derivatives
Trained by NVIDIA. The card and README state that N1.7 uses Cosmos-Reason2-2B, an NVIDIA model with the Qwen3-VL architecture, as its vision-language backbone; Hugging Face metadata lists Qwen/Qwen3-VL-2B-Instruct (developed by the Qwen team at Alibaba Cloud) as the base model of Cosmos-Reason2-2B.1378
- Derived from: Cosmos-Reason2-2B (external site: huggingface.co) — NVIDIA vision-language backbone; no separate catalog record.
- Derived from: Qwen3-VL-2B-Instruct (external site: github.com) — Base model of Cosmos-Reason2-2B per Hugging Face metadata.
Other releases in the NVIDIA Isaac GR00T N family
No other releases in this family have been assessed.
U.S. eligibility
Eligible · basis: U.S. headquarters
The model card names NVIDIA as the model developer. NVIDIA's principal executive offices are in Santa Clara, California, per its Form 10-Q for the quarter ended July 26, 2026. Its vision-language backbone derives from Qwen3-VL-2B-Instruct, recorded under provenance; that base is not U.S.-developed.1978
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.