E1 · SmolVLA
10 minOpen huggingface.co/lerobot/smolvla_base and fill out the sheet.
Done when: the weights license field says apache-2.0 and you noted the base model.
Lesson 3 · 16 min
The model map in 2026
There are dozens of VLAs, and the best one on the chart may be the one you can't use. Before size and benchmark scores, three questions decide: are the weights downloadable, does the license allow your use case, and is access free, paid, or partners only.
Start the lesson
In this lesson you will
- 1Which models are open and which are closed as of Oct 2026
- 2The five ways to get access: Apache, custom license, non-commercial, API, and early access
- 3Where to find the license, size, and robots on a model card
- 4Why LIBERO no longer works for choosing a model
Goal: choose a model for a real use case by license, size, and access (open, non-commercial, API, early access).
01Open × closed
Click each group. The order goes from most free to most restricted.
On Hugging Face, the license sits at the top of the model card (the license field). Gated means you have to request access.
Open doesn't mean free: read the weights' license, which can differ from the code's license.
02How to choose in four questions
Click each step and answer for your own case.
1. Can I download it?
If the model only exists through an API or for partners, you depend on another company for every test.
Watch out for gated access: GR00T N1.7 downloads on its own, but it loads the nvidia/Cosmos-Reason2-2B backbone, which is gated. If you don't request access on the Hub, the run fails with GatedRepoError, according to the Isaac-GR00T README.
2. Can I use it in my case?
Study and research: almost anything that is open works, including G0.5.
Product or client: Apache-2.0 or a license that allows commercial use, read to the end.
Case
An integrator wants to put an arm on a packaging line: G0.5 is out because of its license; GR00T N1.7 or π0.5 make the list.
3. Does it fit my hardware?
SmolVLA has 450M parameters and was built for consumer hardware. GR00T N1.7 has 3B and was built to run on NVIDIA GPUs (check the model card for which ones). DreamZero has 14B.
The LeRobot 0.6 hardware guide measures video memory per policy family: check it before you buy.
4. Does it work on my robot?
Check which robots the model was validated on. N1.7 was validated on Unitree G1, YAM, and AGIBot Genie 1; SmolVLA, on the SO-100/101.
Don't choose by LIBERO: in 2026 several models pass 97% (X-VLA 98.1%; MolmoAct 2-Think 98.1%). Look at RoboArena, MolmoBot, RoboTwin 2.0, or third-party evaluation.
03The model card sheet
Copy it and fill out one per model.
ficha-modelo.md
# Sheet: <model name> Model card URL: Weights license (license field): Code license (repository): Commercial use allowed? (yes / no / read the contract): Gated? (need to request access?): Parameters: Robots it was validated on: Training data (what the card says): How to train or fine-tune (command from the card):
Example from the repository
- lerobot/smolvla_base: license
apache-2.0; baseHuggingFaceTB/SmolVLM2-500M-Video-Instruct; not gated; fine-tune withlerobot-train --policy.path=lerobot/smolvla_base. - nvidia/GR00T-N1.7-3B: weights under the NVIDIA Open Model License Agreement; code on GitHub under Apache 2.0; 3,000,000,000 parameters; training data "Hybrid: Human, Robot, Simulated", 13 datasets; the Cosmos-Reason2-2B backbone is gated.
- Copying the code license: in GR00T the code is Apache 2.0, but the weights are not.
- Assuming open = commercial: G0.5 is open and non-commercial.
- Taking the parameter count from an aggregator: use the card or the paper; the "5.4B" for MolmoAct 2, for example, only shows up on aggregators.
04Practice: read two model cards
You don't need a Hugging Face account to read them.
E2 · GR00T N1.7
15 minOpen huggingface.co/nvidia/GR00T-N1.7-3B and the Isaac-GR00T README and fill out the sheet.
Done when: the sheet separates the weights license (NVIDIA Open Model License) from the code license (Apache 2.0) and marks the backbone as gated.
E3 · Pick for a case
10 minWrite one case of your own in a single line (study at home, prototype for a client, research). Say which of the two models you would test first and why.
Done when: the justification cites license, size, and access, not just benchmarks.
05Test what you learned
The explanation appears after your answer.
Question 1 of 4
A company wants to sell a packaging robot. Which model can it not use without a separate contract?skill
Correct
The G0.5 Community License allows non-commercial use only. A product requires a separate commercial license. The other three are Apache-2.0.Not this one
The G0.5 Community License allows non-commercial use only. A product requires a separate commercial license. The other three are Apache-2.0.Review: Open × closedQuestion 2 of 4
In GR00T N1.7, what is the weights license?
Correct
The code is under Apache 2.0, but the weights are under the NVIDIA Open Model License. This is the most common mistake: copying the repository's license onto the weights.Not this one
The code is under Apache 2.0, but the weights are under the NVIDIA Open Model License. This is the most common mistake: copying the repository's license onto the weights.Review: The model card sheetQuestion 3 of 4
Why isn't LIBERO enough to choose a VLA in 2026?
Correct
A saturated benchmark doesn't separate the models. That's why harder versions (LIBERO-plus, RoboTwin 2.0) and real-world evaluations like RoboArena appeared.Not this one
A saturated benchmark doesn't separate the models. That's why harder versions (LIBERO-plus, RoboTwin 2.0) and real-world evaluations like RoboArena appeared.Review: How to choose in four questionsQuestion 4 of 4
What does the Pi API beta (10/13/2026) offer?
Correct
It's commercial access through an API only. The post says "our latest models" without saying which one. You don't download anything.Not this one
It's commercial access through an API only. The post says "our latest models" without saying which one. You don't download anything.Review: Open × closedQuiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as complete.Pick an option. The explanation appears after your attempt.
06Summary
Three questions before the chartCan I download it, can I use it for my case, does it fit my hardware.
Code ≠ weightsThe weights license may be different. Gated means you have to request access on the Hub.
LIBERO saturatedAbove 97%, everyone ties. Use a real-world evaluation or a harder benchmark.