Stanford / UC Berkeley / TRI et al. (OpenVLA team)
OpenVLA
7.54B Commercial terms unverified 7/8 verified
1 of 8 required fields are unverified. They are shown as badges, not filled with defaults. Sources for verified fields are listed below.
Facts
- Released
- 2024-06 1 Source
- Parameters
- 7.54 B 1 Source
- Architecture
- Monolithic autoregressive 0 Source
- License
- MIT (tagged on code and weights repositories); weights derived from Llama-2, subject to the Llama 2 Community License 2 Source
- Commercial
- Commercial terms unverified 0 Source
- Weights
- huggingface.co ↗ 1 Source
- Code
- github.com ↗ 1 Source
- Paper
- arxiv.org ↗ 1 Source
- Input modalities
- rgb · language 0 Source
- Action space
- 7-DoF end-effector pose delta + gripper (discrete tokens) 0 Source
- Cross-embodiment
- Pretrained on approx. 970k Open X-Embodiment trajectories 1 Source
- Pretraining data (hours)
- Unverified 0 Source
- Target forms
- Single arm 0 Source
Architecture · Prismatic VLM: DINOv2 + SigLIP dual vision encoders + Llama-2 7B; actions discretised into tokens, output autoregressively
Sources · 3 sources
license,release_date,code_url https://github.com/openvla/openvla Official page · Fetched 2026-09-04 · No snapshot stored yet README: All code is made available under the MIT License; Models derive from Llama-2, subject to the Llama Community License; first released 2024-06-13
params_b,weights_url,license https://huggingface.co/api/models/openvla/openvla-7b Model hub page · Fetched 2026-09-04 · No snapshot stored yet safetensors 7,541,237,184; model card license: mit; createdAt 2024-06-10
paper_url,cross_embodiment https://arxiv.org/abs/2406.09246 Paper / official docs · Fetched 2026-09-04 · No snapshot stored yet OpenVLA: An Open-Source Vision-Language-Action Model；970K trajectories from Open X-Embodiment
Activity signals
Public counters fetched daily from GitHub, Hugging Face and ModelScope. Numbers only; no judgement. · fetched 2026-10-18
GitHub stars
7,169
+130 / 30 d
Commits · 90 d
0
last push 575 d ago
HF downloads · 30 d
153,562
all-time 18,113,297 · 262 likes
Weights last modified
2026-02-17
244 d ago
ModelScopezixiaoBios/openvla-7b · 3,188 downloads · updated 2025-03-05 · same-name mirror, not the vendor org
Hugging Faceopenvla/openvla-7b
stars · last 30 d
Public benchmark scores
As printed in the source; protocols differ between papers. Full table with notes on the benchmarks page.
| Benchmark | Subset | Score | Source |
|---|---|---|---|
| LIBERO | avg | 76.5 | paper · Table 12, row 'OpenVLA fine-tuned (ours)', Average SR (76.5 ± 0.6%) |
| LIBEROOpenVLA-OFT | avg | 97.1 | paper · Table I, row 'OpenVLA-OFT' (multi-input, modified dataset), Average |
| LIBERO | goal | 79.2 | paper · Table 12, row 'OpenVLA fine-tuned (ours)', LIBERO-Goal SR (79.2 ± 1.0%) |
| LIBEROOpenVLA-OFT | goal | 97.9 | paper · Table I, row 'OpenVLA-OFT' (multi-input, modified dataset), Goal |
| LIBERO | long | 53.7 | paper · Table 12, row 'OpenVLA fine-tuned (ours)', LIBERO-Long SR (53.7 ± 1.3%) |
| LIBEROOpenVLA-OFT | long | 94.5 | paper · Table I, row 'OpenVLA-OFT' (multi-input, modified dataset), Long |
| LIBERO | object | 88.4 | paper · Table 12, row 'OpenVLA fine-tuned (ours)', LIBERO-Object SR (88.4 ± 0.8%) |
| LIBEROOpenVLA-OFT | object | 98.4 | paper · Table I, row 'OpenVLA-OFT' (multi-input, modified dataset), Object |
| LIBERO | spatial | 84.7 | paper · Table 12 (Appendix E.2), row 'OpenVLA fine-tuned (ours)', LIBERO-Spatial SR (84.7 ± 0.9%) |
| LIBEROOpenVLA-OFT | spatial | 97.6 | paper · Table I, row 'OpenVLA-OFT' (multi-input setup, modified dataset), Spatial |
| Real-robot evaluation (self-reported) | bridgedata_v2_widowx_17_tasks | 70.6 | paper · Table 4 (Appendix B.1.3), bottom row 'Mean Success Rate', column 'OpenVLA (ours)' (70.6±3.2%) |
| Real-robot evaluation (self-reported) | franka_droid_finetuned_avg | 58.3 | paper · Table 7, Franka-DROID (15Hz) 'Average' row, column 'OpenVLA (ours)' (58.3±7.2%) |
| Real-robot evaluation (self-reported) | franka_tabletop_finetuned_avg | 67.2 | paper · Table 7 (Appendix B.3.2), Franka-Tabletop (5Hz) 'Average' row, column 'OpenVLA (ours)' (67.2±4.0%) |
| Real-robot evaluation (self-reported) | google_robot_12_tasks | 85 | paper · Table 6 (Appendix B.2.2), bottom row 'Mean Success Rate', column 'OpenVLA (ours)' (85.0±4.6%) |
Deployment recipe
Environment, weights, run command and the pitfalls we hit, exactly as executed on rented GPUs (venv openvla, 15.2 GB of weights, 2 pitfalls).
Open recipe →Record history
2026-09-04
Sources fetched
license,release_date,code_url · params_b,weights_url,license · paper_url,cross_embodiment
Latency & cost
cost = p50 ms ÷ 1000 × rental ÷ 3600 × 1000
| Hardware | p50 ms | p95 ms | Peak VRAM | Cost per 1k inferences ¥ | samples |
|---|---|---|---|---|---|
NVIDIA RTX 4090 ¥1.68/h · GongJi Compute (2026) | 162 | 173 | 14.41 GB | 0.075 | 1 Preliminary (n<3) Maintainer measured |
NVIDIA RTX 5090 ¥2.4/h · Runyun 2.4 / GongJi 2.5 / Paratera Cloud 2.98 (2026) | 146 | 150 | 14.41 GB | 0.098 | 1 Preliminary (n<3) Maintainer measured |
NVIDIA A100 80GB no rental price | — | — | — | — no latency data | No data |
NVIDIA Jetson AGX Thor no rental price | — | — | — | — no latency data | No data |
NVIDIA H100 80GB no rental price | — | — | — | — no latency data | No data |
NVIDIA A800 80GB no rental price | 187 | 202 | 14.41 GB | — no rental price for this hardware | 1 Preliminary (n<3) Maintainer measured |
Compatibility
Colored only where we hold evidence. An amber dot means the vendor declares this form as a target; that is a claim, not a verified compatibility.
| Unitree G1 | Unitree H1 | AgiBot Genie G1 | AgileX Piper arm | SO-101 | Franka FR3 / Panda | Fourier GR-2 | WidowX 250 S | Google Robot（Everyday Robots） | ALOHA 2 / Cobot Magic (dual-arm) | Galaxea R1 Pro | Franka Emika Panda | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OpenVLA |
Legend Officially supportedCommunity reproducedFine-tunable in theoryUnsupportedUnknown Vendor-declared target form 3 / 12 cells with a definite status
Reproduction records
Did this model run on your robot? Report it with a link (log, video or repo). Records are reviewed by the maintainers; a verified “works” or “works after fine-tuning” upgrades the matrix cell to “community verified” with your link as evidence. Anonymous is fine; nothing but the fields below is stored.
No verified record yet for this model.
Clear direction, grounded judgement.