Stanford

OpenVLA↗

OpenVLA is a VLA from Stanford, released June 13, 2024. It has no overall rank on the Robot Intelligence Index, which needs both a simulation and a real-robot result. Its best result is #719 of 823 on Paper-reported. 7B open VLA (Llama 2 + DINOv2/SigLIP) trained on 970k Open X-Embodiment episodes; discretized action tokens.

Release date: Jun 13, 2024

Developer
Stanford
Type
VLA
Model weights
Open
License
MIT
Parameters
7.5B
Pricing
Self-hosted
Benchmarks
3
Overall rank
Not ranked
Needs a real-robot result to rank overall.
Real robots rank
Not ranked
No real-robot results yet.
Simulation rank
#49of 49
Rating 952 · range 36–49
Rating
1087± 4
10601989
Cost / task · est.
$0.012
$0.0087$17.74
Time / task · est.
14s
9.5 s1 h 22 min
Rank on each benchmark
BenchmarkScoreRank
Paper-reportedLIBERO avg76.5719 / 823LIBERO-Pro2Unperturbed35.119 / 20LIBERO-PlusOverall15.611 / 11

Per task attempt; the tiles show the median across benchmarks. Self-hosted compute is priced at one H100 at $3.29 per hour (Lambda on-demand, Oct 2026).

Other models from Stanford

Inference

Single action per forward pass: 239.6 ms / 4.2 Hz on A100 (LIBERO, 1 image); 543 ms / 1.8 Hz with 3 images (ALOHA). Independent: 276 ms on RTX 4090, 173 ms on H100.