smolvla
Supported450M params
Overview
vla.cpp's "golden reference" architecture — every other arch is validated against this recipe. Simplest setup: no gating, no extra stats files. Start here.
Architecture
- Vision backbone
- SigLIP-So400m
- Language backbone
- SmolLM2-360M
- Action head
- flow-matching cross-attention expert
- Action steps per chunk
- 1
- Solver steps
- 10
- Camera views
- 2 — observation.images.image, observation.images.image2
Hardware fit
Fits the 8GB Orin Nano target
Benchmarks
| Hardware | Success rate | Step time | Inference time | Memory |
|---|---|---|---|---|
| RTX 3060 | 90.5% | 28.16 ms | 54.8 ms | 1410 MiB VRAM |
| Orin Nano 8GB | — | 141.81 ms | — | 2031 MiB RSS |