Supported Models
OpenVLA-7B
7B-parameter VLA using SigLIP + DinoV2 fused vision encoder with Llama 2 backbone. Predicts discrete action tokens.
| Spec | Value |
|---|---|
| Parameters | 7B |
| Min GPU | A100 40GB |
| VRAM | ~28 GB |
| Inference | ~8 Hz |
| License | Apache 2.0 |
model = VLAModel.from_preset("openvla-7b")
SmolVLA-450M
Compact 450M VLA by Hugging Face using flow-matching. Runs on consumer GPUs.
| Spec | Value |
|---|---|
| Parameters | 450M |
| Min GPU | RTX 3090 |
| VRAM | ~8 GB |
| Inference | ~25 Hz |
| License | Apache 2.0 |
model = VLAModel.from_preset("smolvla-450m")
Dream-VLA-7B
7B VLA built on diffusion language model backbone with parallel action generation.
| Spec | Value |
|---|---|
| Parameters | 7B |
| Min GPU | A100 40GB |
| VRAM | ~30 GB |
| Inference | ~6 Hz |
| License | Apache 2.0 |
model = VLAModel.from_preset("dream-vla-7b")
Custom Models
Register your own model:
from vlarobot.models.registry import MODEL_REGISTRY
MODEL_REGISTRY["my-model"] = ("my_package:MyModel", "org/model-id")