Gated models

Most models are open-weight and need no token. A few are gated: their weights sit behind an upstream license that you accept with your own Hugging Face account. oh-my-mlip never redistributes gated weights — they are downloaded on your machine with your token, after you accept the license.

Which models are gated

Models	License page
all UMA variants	https://huggingface.co/facebook/UMA
eSEN-30M-OAM	https://huggingface.co/facebook/OMAT24

Everything else is open. The registry marks each model: gated is true and license_url points at the page to accept (see Supported models).

Ask your LLM

Install UMA and check it runs on my GPU.

Your agent notices the model is gated, shows you the license page to accept, and asks you to log in to Hugging Face in your own terminal. It never asks you to paste a token into the chat.

The flow (once per model and machine)

1. Accept the license on the model's license page, logged in with the account whose token you will use.
2. Log in so the token is available: hf auth login (details and alternatives: Hugging Face token). source env.sh does not set a token for you, so no token ends up in the repo.
3. Run normally. The first call downloads the weights into the framework's own cache; later runs reuse it.

Run it yourself

source env.sh
hf auth login
python run_examples/single_point.py UMA --version UMA-s-1p2-OMAT

Without a token or an accepted license

The download fails on purpose: nothing retries or falls back to a mirror. An HTTP 401 or 403 response means the license is not accepted yet, or the token belongs to a different account. Accept the license, log in, and run again.

Policy

• Gated weights are never committed to this repository or uploaded anywhere by oh-my-mlip.
• They are always downloaded on demand with the user's own credentials.