← Leaderboard

OpenVLA-OFT

VLA
Stanford · Feb 27, 2025 · built on OpenVLA

Optimized fine-tuning recipe for OpenVLA, adding parallel decoding, action chunking and continuous actions. Several RL post-training methods use it as their base.

Type
VLA
Parameters
7B
Backbone
Prismatic (Llama 2 7B)
Weights
Open
License
MIT
Commercial use
Allowed
Released
Feb 27, 2025
Robots
Franka Panda (sim), ALOHA
SO-101
–
LeRobot
–

Benchmark results

3 of 13 benchmarks
#40 of 400.0%, self-reported by the authors, not a leaderboard entry
#19 of 2471.4%, self-reported by the authors, not a leaderboard entry
LIBEROSaturatedSimulation
#31 of 4697.1%, self-reported by the authors, not a leaderboard entry
Not reported on 10 benchmarks
Robustness drop: 97.1% on LIBERO → 71.4% on LIBERO-Plus. The perturbed variants show how much of the LIBERO score is memorization.

Adoption

updated Nov 17, 2026
HF downloads · 30d
11.5K
HF downloads · all time
134.1K
HF likes
15
GitHub stars
1.4K
GitHub forks
213

Running the model

HardwarePrecisionConfigLatency / chunkStepsVRAMChunkControlSource
A100–parallel decoding + chunking + L1 regression, 1 image (LIBERO)72.9 ms––8–Feb 27, 2025 ↗

Trained on

Runs on

Wrong or missing number? Report an error

Built on OpenVLA-OFT