Model library
Every model, and the ones we will not serve
Everything we serve, plus the ones we refuse and the clause that stops us.
Every licence here was read from the repository file rather than the metadata field —
the most downloaded robot model on the Hub declares no licence at all, and the terms
that matter live only in the file.
Vision-Language-Action VLAPOST /v1/act
Feed-forward models that read the latest frame from each camera and return a chunk of joint commands. One request body serves all of them, whatever width they answer in.
World-Action Model WAMPOST /v1/act
These imagine the next seconds before committing to a move, so they read a window of frames rather than one. Same endpoint, same body — you simply send more frames. One GPU carries under a single robot, which is why they price differently.
Vision-Language Model VLMPOST /v1/chat
The reasoning half. Messages in, a message out, in the same envelope the action endpoint answers with. Some are self-hostable, which is where the long-run cost floor comes from.
Refused on licencelisted, not served
Kept visible rather than deleted. Remove them and the next reader rediscovers the model, adds it, and learns the licence the expensive way. Calling one returns 403 with the clause that blocks it.
refusedCosmos Policy ALOHA Predict 2
cosmos-policy-aloha-predict2-2b
NSCLv1 — only NVIDIA may use it commercially. ALOHA 93.6% is the best real-robot score in the family and we still cannot serve it.
403 at the gate
refused
DreamZero DROID
dreamzero-droid
CC-BY-NC-4.0 — the paper's headline checkpoint. We serve the Apache-2.0 AgiBot build instead.
403 at the gate
refused
GROOT N1
groot-n1-2b
The 2025 original: non-commercial. N1.7 changed licence in 2026 and is served above — two versions, opposite terms.
403 at the gate
vla
vla
vla
vla
vlm
vlm
vlm
vlm