Model survey · June 2026
The VLA
model zoo, 2026
Ten vision-language-action models that turn pixels and a language goal into motor actions. This is the concrete architecture behind the System 1 / System 2 and generalist theses. Filter by weights and action method, sort by size or year.
The thesis
A VLA takes in pixels and a goal and emits actions. How it produces those actions matters more than its parameter count.
This survey maps ten models that defined the field from 2023 to 2025. Three choices separate them: how actions are generated (autoregressive tokens, diffusion, or flow matching), whether the weights are open or closed, and whether intelligence is one network or an explicit fast/slow pair. Parameter counts are totals unless a split is noted.
The field
Ten ways to turn seeing into doing.
Filter by weights and action method, or search developers and embodiments. Sort by model, developer, release, or parameters.
| Weights | Action method | Embodiment |
|---|
My synthesis
Three movements, visible when you sort.
Sort by release and the trajectory appears. Action generation moved from text tokens to diffusion and flow matching. Useful size fell from RT-2's 55B to SmolVLA's 450M. And intelligence split into two: a slow VLM planner driving a fast action expert, the System 1 / System 2 design from note 002 made into product.
Strategic implication
The base model is shared. Advantage comes from what you feed it.
A VLA is the generalist of note 005 made concrete, and the open ones, including π0, GR00T, OpenVLA, and SmolVLA, mean a robotics company rarely starts from zero. The differentiator moves up the stack.
It becomes the proprietary demonstrations used to post-train the generalist, the embodiment-specific System 1 the open VLM cannot supply, and the data engine that turns each deployment into the next training set.
A VLA turns pixels + a goal into actions. The frontier moved from tokens to flow, and from one model to two.