# SACQ Shrinks VLA Model Memory 4.83× at 1.5 Bits

> A research team reports that its quantization method preserved robot-task success with much smaller model weights, tested on GR00T N1.7 and LIBERO.

Oossa · 2026-11-02 · https://oossa.com/en/sacq-shrinks-vla-model-memory-4-83-at-1-5-bits

Researchers introduced SACQ, a method for compressing the weights of vision-language-action models used to control robots. It assigns precision based partly on how compressing each layer affects the model’s final action.

In tests with GR00T N1.7, the team reports 93.85% average success across four LIBERO suites using 1.5 bits per language/action weight, while cutting model memory by 4.83 times. These are results reported in the arXiv paper; the supplied abstract does not establish independent validation or identify deployment hardware.

## The facts

- GR00T N1.7 achieved 93.85% average success across four LIBERO suites at 1.5 bits per language/action weight, according to the researchers.
- The researchers report a 4.83× reduction in model memory.

## Why it matters

For teams trying to run robot-control models on memory-limited platforms, the reported compression could make deployment easier, but the abstract does not say which hardware can run SACQ or confirm the results independently.

## Sources & references

1. [SACQ: Structure-Aware Codebook Quantization for Ultra-Low-Bit Vision-Language-Action Models](https://arxiv.org/abs/2610.41408) – arXiv cs.CV, 2026-11-02

Last updated: 2026-11-02