GR00T N1 vs N1.5 vs N1.7: the version chooser, answered

Pick N1.7 for anything new. It is NVIDIA's current release, it replaces N1.5 everywhere NVIDIA now points developers, and it carries the best published results in the family. Pick N1.5 only to finish or repeat an experiment that already started on it. Pick N1 only to read history. And if your robot must ship as a product, no version clears that bar: the license file inside every GR00T release, N1.7 included, permits research and evaluation only. That answer does not change by version, whatever the launch posts say.

The verdict rests on publisher sources read on October 26, 2026. Sources: model cards, the license files themselves, one research page per release, and the GitHub repo. Where NVIDIA published no number, the table cell gets a dash. There is a lot of that. Knowing it is half the choice.

What people actually ask

Two demand signals shape this query. First, search completions pair the versions against each other. The strings "gr00t n1 vs n1.5", "gr00t n1.5 vs n1.7", "gr00t n1 vs n1.7", and "which gr00t version" all completed live in the October 2026 search census. So did the fine-tuning string "gr00t n1 5 fine tuning", which still completed when re-checked on October 26, 2026. Second, competitors wrote pages, but each covers a slice. The LearnOpenCV explainer covers N1 versus N1.5. It dates to June 2025 and stops there. The AY-Robots page compares N1.7 against N1.5 for its own training form. The Particula agency post, from August 2026, compares GR00T against three other model families. It never splits the versions apart.

Five real questions sit under all of that. Which version do I start a new robot on. Is my N1.5 project now obsolete. Can I ship any of them in a product. Which version fits the hardware I own. And what happened in between. The rest of this page answers each one in order.

The chooser in one table

Read the table as a decision aid, not a ranking. The last row is the one people miss. A version choice cannot fix a license problem. The license is the same file in all three.

Three releases, one family

GR00T N1 arrived on March 17, 2025. NVIDIA called it the first open foundation model for humanoid robot skills. The white paper came the next day. It describes the design every later version keeps. A vision-language module reads cameras and text. A diffusion head turns that reading into motor actions. Both train end to end. The released checkpoint, nvidia/GR00T-N1-2B on Hugging Face, counts 2.19 billion parameters. The card's own release fields carry the date. NVIDIA showed it on the Fourier GR-1 humanoid, doing two-handed tasks from little data.

GR00T N1.5 followed on June 11, 2025. The NVIDIA research page is titled "An Improved Open Foundation Model for Generalist Humanoid Robots". The vision-language backbone was frozen and rebuilt from Eagle 2.5, tuned for grounding. A world-modeling goal called FLARE joined training. The data mix grew: Open X-Embodiment sets, DreamGen neural trajectories, AgiBot recordings. The checkpoint grew to 2.72 billion parameters. NVIDIA also posted a community fine-tune tutorial for the SO-101 arm the same day. That is why so much hobby material still points at N1.5.

GR00T N1.6 shipped on December 15, 2025. Almost nobody wrote a guide about it. Its research page records the real changes. A Cosmos-2B vision-language backbone, with native aspect ratio support. A diffusion head doubled to 32 layers. State-relative action chunks instead of absolute targets. Thousands of hours of new teleoperated data, from bimanual YAM arms, AgiBot Genie-1, and Unitree G1. The checkpoint counts 3.29 billion parameters. It matters to the chooser for one reason. It is the direct predecessor N1.7 gets measured against.

GR00T N1.7 was announced in early access on April 17, 2026. The launch post describes a 3 billion parameter reasoning model on a Cosmos-Reason2-2B backbone. It uses a relative end-effector action space shared by human and robot bodies. It was pretrained on 20,854 hours of human egocentric video under the EgoScale program. The GitHub repo now carries a General Availability banner, a LeRobot integration with its own policy type, and a fine-tune example for the SO-100 arm. Its own SO-101 support is one config file inside an example. The fuller path runs through LeRobot. On July 7, 2026, NVIDIA wrote that GR00T 1.7 replaces N1.5 in LeRobot. That post also covers the Isaac Teleop data-capture tool, which lives in the LeRobot path rather than the model repo. Developers should move, because N1.5 is no longer supported there.

The specs side by side

The download counts come from the Hugging Face API on October 26, 2026. The market has voted. N1.7 pulls roughly 46 times the monthly downloads of N1.5. It pulls 170 times N1. N1.5 still beats N1 by under four times. That fits its role: the version older experiments run on. GR00T N1.6 sits between at 10,048 downloads a month, with almost no third-party coverage.

What the published numbers say

NVIDIA published exactly one full cross-version table. It covers N1 against N1.5. It sits on the N1.5 research page, dated June 11, 2025.

Every N1.7 cell is a dash on purpose. No N1.5 versus N1.7 table exists. NVIDIA never published one. The official line is one sentence in the repo readme. N1.7 "delivers comparable performance to N1.6, with improved generalization and language-following". The N1.6 research page says N1.6 beats N1.5 on its benchmarks, and on real YAM, Genie-1, and G1 robots. Again without one shared table. Chain the two sentences. Each release beat the one before it, on that release's own tests, under conditions that moved each time.

So the version choice is not a benchmark choice. You cannot look up an N1.5 number and an N1.7 number for the same task from NVIDIA today. If a blog hands you such a pair, one side is not a publisher number. Run your own test on your own task. The evaluation cost guide budgets that work.

The license answer, version by version

This is the part the version numbering hides. On October 26, 2026 I fetched the LICENSE file from each base checkpoint repo and compared them. The N1 file is 4,062 bytes. The N1.5, N1.6, and N1.7 files are 4,063 bytes each. The text is identical apart from a trailing newline. Every version ships the same "NVIDIA License". Its section 3.3 reads:

3.3 Use Limitation. The Work and any derivative works thereof only may be used or intended for use non-commercially. Notwithstanding the foregoing, NVIDIA Corporation and its affiliates may use the Work and any derivative works commercially. As used herein, "non-commercially" means for research or evaluation purposes only.

Three sentences, three rules. The weights are for research and evaluation. NVIDIA and its affiliates keep commercial rights. Fine-tunes inherit the limit.

What differs across versions is the claim layer around the file. The N1 card links a non-commercial license PDF. It says nothing about commercial use. The N1.5 card states plainly that the model is for non-commercial use. The N1.7 model card goes the other way. It claims the model is ready for commercial and non-commercial use, and names the NVIDIA Open Model License. The repo readme calls N1.7 "fully commercially licensable under Apache 2.0". The April 2026 launch post and forum announcement say commercial licensing "is now included", and frame that as the milestone of the release.

The file did not change. It still has not changed as of today. Three open discussions on the N1.7 repo ask NVIDIA which paper governs. They opened April 21, July 28, and August 25, 2026. None has a reply from NVIDIA. The pattern was set earlier, in a July 2025 forum answer. NVIDIA staff wrote that the weights carry a research-only license due to pre-training data limits. They called it a data issue, not a license choice. They said no permission process exists.

Which paper governs your copy is not a mystery. The license file that ships with the weights. It travels with every download. Its own terms bar you from passing the weights on under friendlier terms. When the readme and the file disagree, a careful reader reads the file first. This page is engineering guidance, not legal advice. For a shipping product, read the file yourself and put the question to your lawyer. The deeper walkthrough, with all four NVIDIA claims side by side and the Apache 2.0 way out, is the GR00T license guide.

Hardware: what each version needs

The current repo states one floor for N1.7. Inference runs on one GPU with 16 GB of memory or more. Fine-tuning wants 40 GB or more per GPU. H100 and L40 are the named preferred parts. A 24 GB gaming card therefore serves any version for inference. Fine-tuning is where the versions split.

The N1.5 era recipe was friendlier to owned hardware. Its branch readme lists H100, L40, RTX 4090, and A6000 as tested parts. It records that two RTX 4090 cards, or two A6000 cards, ran LoRA fine-tuning. It notes a single 4090 needs one extra flag to avoid running out of memory. The community built on that. The Seeed wiki walks a full N1.5 fine-tune on the SO-101 arm, then deploy to Jetson Thor. It is far from the only such tutorial. N1.7 fine-tune tutorials exist too. The main branch now ships an SO-100 example, and the SO-101 route runs through LeRobot. But the official memory ask moved up, not down.

Published control speeds, with the conditions stated:

The two Thor rows for N1.7 are not a typo. The readme and the model card used different setups, and both printed their conditions. That spread is normal in this family. Every speed claim carries its setup. Comparing across setups is how readers get burned. The deeper Jetson walkthrough, Orin limits included, is the GR00T on Jetson guide.

Nobody serves any version for you

As of October 26, 2026, no host we checked serves any GR00T version behind an API. The Hugging Face inference provider mapping is empty on all four base repos. A fal search returns zero GR00T models. Replicate's NVIDIA page lists none. DeepInfra lists none. NVIDIA's own build portal hosts one GR00T-branded item. It is not a policy. It is a blueprint that generates synthetic motion data from human demos. A data tool for training. Not a model you can call for actions.

That gap holds across the whole family, and it is older than N1.7. The big hosts survey covers the pattern across robot models. One nuance. A third-party service will fine-tune GR00T for you on rented cards. AY-Robots is the example in these results, and their verdict page, dated August 2026, tells customers to pick N1.7 for anything new. A training service is not a hosted policy. After training, the weights come back to you. Running them is still your hardware problem.

So the honest chooser has a hosting row too. Run it yourself, on a workstation card or a Jetson module. Or do not run it. The Jetson versus cloud guide works that trade for robot budgets.

If your robot must ship

The license file is why the chooser cannot end at "N1.7". A robot built on weights you cannot license for sale is a robot you cannot sell. Two model families do the same class of work under a clean Apache 2.0 grant on the weights. Both were named in the license guide above.

The first is pi05 from the openpi project. About 3.6 billion parameters, with released tunes for standard arms, including a DROID tune. The second is MolmoAct2 from the Allen Institute for AI. It is tuned natively for the SO-100 and SO-101 arms that most GR00T tutorials also target. Both take the same request shape GR00T takes. Camera frames, an instruction, and the joint state go in. An action chunk comes out.

This platform keeps written records for both families on its action route, metered per call. As of October 26, 2026 neither row is on the public list, and this page quotes no live rate. The models index is the live catalog. Rates for listed rows render on the pricing guide. The per-checkpoint pages carry the request shapes: pi05 and MolmoAct2. If self-hosting suits you better, the local versus hosted break-even guide runs the full math on both sides.

Choose in 60 seconds

Work down the list. Stop at the first line that fits your case.

1. Starting fresh, on any hardware from a 4090 up: N1.7.
2. Reproducing a paper, or continuing a run from 2025: the version the run used. N1.5 most likely.
3. Wanting the deepest tutorial trail and the lowest official fine-tune floor: N1.5, knowing it is superseded.
4. Deploying to Jetson Thor or Orin: N1.7. Read its deploy pages first.
5. Selling what you build: none of them. Take the Apache 2.0 route and keep your records.

Common questions

Which GR00T version is the newest? N1.7. It was announced in early access on April 17, 2026. The GitHub repo now presents it as the General Availability release. N1 and N1.5 are superseded. NVIDIA's July 2026 LeRobot post says to move off N1.5.

Is GR00T N1.5 still supported? Not in LeRobot. NVIDIA's July 7, 2026 post says GR00T 1.7 replaces N1.5 there, and N1.5 is no longer supported. The checkpoint stays downloadable for repeating older work.

Can I use GR00T N1.7 commercially? The license file shipped with the N1.7 weights says no. Section 3.3 limits use to research and evaluation, and keeps commercial rights with NVIDIA and its affiliates. The model card, the readme, and the launch post all claim commercial rights. Repo discussions asking about the conflict sit open, with no NVIDIA reply. One paper ships with your copy: the file. The full analysis is the license guide.

What is GR00T N1.6? The December 15, 2025 release between N1.5 and N1.7. It moved the backbone to Cosmos-2B, doubled the action head, and switched to state-relative actions. Its weights are public under the same license file. NVIDIA calls N1.7 comparable to it in performance, with better generalization. Most choosers skip it.

Which version should I fine-tune on an SO-100 or SO-101 arm? N1.7 for new work. The model repo ships an SO-100 example. The SO-101 path runs through LeRobot, which carries a native GR00T policy type plus the Isaac Teleop tool for collecting data. N1.5 keeps the deepest tutorial trail, including the Seeed wiki path to Jetson Thor. One dated caution: community reports from the early-access period said the example data assumed the SO-101 joint layout. Today's repo example data is SO-100, so check the joint layout of any dataset before training.

Can I run GR00T on one RTX 4090? For inference, yes. The stated floor is 16 GB of GPU memory. For fine-tuning, the N1.7 recipe asks for 40 GB or more per GPU. The N1.5 recipe documents LoRA on a pair of 24 GB cards, plus a single-card flag. If you hit memory errors on other families' checkpoints, the OpenVLA-OFT memory guide covers the general fixes.

The short version

The ladder is short. N1 proved the idea in March 2025. N1.5 made it follow language, and the June 2025 numbers doubled its success rates. N1.6 rebuilt the backbone in December 2025. N1.7 is the current model: human-video pretraining, relative actions, a real deploy path, and every new tutorial written against it. Choose it for anything new. Keep N1.5 for old runs, and N1 for archaeology. Before any of it goes near a product, read the license file in the repo you downloaded. The file, not the version number, decides what you may ship. When you want the runnable path under Apache 2.0 terms, the action route records and the pricing guide are the place to start, and sign up to create a key.

Last verified: 2026-10-26