NASA Puts Google’s Gemma Large Language Model in Orbit

Summary

NASA tested the first vision-language model that analyzes images from a satellite sensor by sending Google's Gemma 3 large language model into space.

Why it matters

This breakthrough enables researchers to interact directly with spaceborne systems via prompts, transforming traditional command protocols. Moreover, the 4-bit compressed model's low power consumption while running on an Nvidia Jetson Orin AGX demonstrates the feasibility of deploying large language models in space.

Highlights

• Gemma 3 was run in a compressed 4-bit format within the NAVI-Orbital agency framework and achieved 88% accuracy on a benchmark dataset of 7,960 images.
• The satellite, powered between 150–500 watts, ran the large language model on an Nvidia Jetson Orin AGX and generated textual summaries and answers to questions.
• Image analysis overcomes limited bandwidth constraints through 'semantic compression,' enabling the near-instant detection of emergencies such as fires.

Critical take

The benchmark results were obtained on the ground, not in space; therefore, the real-time accuracy of the model's performance in orbit has yet to be fully validated.