In the paper at the bottom of googles page, this VLA says it is built on the foundations of Gemini 2.0 (hence my quotations). They'd be using Gemini 2.0 rather than llama.

https://arxiv.org/pdf/2503.20020