← Leaderboard
OpenVLA
VLAStanford / UC Berkeley et al. · Jun 13, 2024
Open-source 7B VLA trained on Open X-Embodiment data that predicts discretized action tokens. It is a reference baseline for the LIBERO family.
- Type
- VLA
- Parameters
- 7B
- Backbone
- Prismatic (Llama 2 7B)
- Weights
- Open
- License
- MIT
- Commercial use
- Allowed
- Released
- Jun 13, 2024
- Robots
- WidowX (Bridge), Google Robot, Franka
- SO-101
- –
- LeRobot
- –
Benchmark results
4 of 13 benchmarksLIBERO-PlusSimulation
- Paper / model card ↗Oct 2025
#24 of 2416.3%, self-reported by the authors, not a leaderboard entry
LIBERO-PROSimulation
- Paper / model card ↗Oct 2025
#5 of 657.5%, self-reported by the authors, not a leaderboard entry
- Paper / model card ↗Jun 2024
#45 of 4676.5%, self-reported by the authors, not a leaderboard entry
- Paper / model card ↗Aug 2025· visual matching
#19 of 191.0%, self-reported by the authors, not a leaderboard entry
Not reported on 9 benchmarks
Robustness drop: 76.5% on LIBERO → 16.3% on LIBERO-Plus → 57.5% on LIBERO-PRO. The perturbed variants show how much of the LIBERO score is memorization.
Adoption
updated Nov 17, 2026Running the model
| Hardware | Precision | Config | Latency / chunk | Steps | VRAM | Chunk | Control | Source |
|---|---|---|---|---|---|---|---|---|
| A100 | – | fine-tuned, autoregressive, 1 image (LIBERO) | 239.6 ms | – | – | 1 | – | Feb 27, 2025 ↗ |
| RTX 4090 | bf16 | no compilation or speculative decoding | 167 ms | – | 15 GB | 1 | – | Jun 13, 2024 ↗ |
| RTX A5000 | int4 | 4-bit quantization | 333 ms | – | 7 GB | 1 | – | Jun 13, 2024 ↗ |
Trained on
Wrong or missing number? Report an error