bitvla

Overview

1-bit/ternary weights, smallest footprint of any served model (1.3 GiB resident on Orin Nano) despite being 2.4B params. Highest success rate of all 7 archs (100% on LIBERO-Object, RTX 3060). No iterative solver loop — single forward pass emits the whole chunk. vla.cpp includes a custom tensor-core ternary GEMM kernel for this model specifically (4-4.6x additional speedup over the naive path).

Architecture

Vision backbone

BitSigLIP-L (ternary)

Language backbone

BitNet-2B (ternary)

Action head

single-pass MLP regression (no solver loop)

Camera views

2 — observation.images.image, observation.images.image2

Hardware fit

Benchmarks