TinyVLA
Compact VLA family that skips large-scale pre-training and still beats OpenVLA at 14ms inference.
About
TinyVLA (September 2024, arXiv 2409.12514; accepted at IEEE RA-L 2025) is the anti-scale argument in VLA form. Instead of pre-training on millions of robot trajectories, it starts from a fast off-the-shelf multimodal backbone and adds a diffusion policy decoder fine-tuned with LoRA on task data only. The biggest variant, TinyVLA-H (1.3B total, 143M trainable), reports 94.0% average success on five real single-arm tasks versus 68.3% for OpenVLA, while running inference in about 14ms on an A6000 — roughly 20x faster than the 7B OpenVLA. It also held up on bimanual UR5 tasks where OpenVLA scored zero. The catch, and it is a real one: I found no public code or weight release on the project page during this research, so right now TinyVLA is a paper with strong numbers, not something you can download.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper (arXiv 2409.12514): TinyVLA-H 94.0% avg on 5 real single-arm tasks (OpenVLA 68.3%); 14ms inference vs ~292ms for OpenVLA (author-reported)
- No independent verification; no public weights to test against
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: Midea/ECNU paper (arXiv 2409.12514, accepted IEEE RA-L 2025) with real-robot single-arm and bimanual evaluation against OpenVLA. Strengths: data-efficiency argument with striking reported speed and success numbers. Docked for: (1) no public code or weights found — nothing independently checkable; (2) all numbers are author-reported. License: none — no public release.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 4.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Midea/ECNU paper arXiv 2409.12514 (Sep 2024), accepted IEEE RA-L 2025; TinyVLA-H 1.3B, 143M trainable
- · 94.0% avg on 5 real single-arm tasks (OpenVLA 68.3%); ~14ms inference vs ~292ms OpenVLA (author-reported)
- · No public code or weights found on the project page at research time