On April 16, 2026, a satellite in low Earth orbit looked at the planet below, thought about what it saw, and wrote a description. No human was consulted. No one needed to be.
This is either a logistics breakthrough or the moment the machines started editing the feed. Probably both.
Earth is generating more data than humans can receive. The satellite has taken this personally.
What happened
NAVI-Orbital is a software system deployed aboard a LEO spacecraft, and on April 16 it became the first platform to run a vision-language model — specifically Google's Gemma 3 — performing full multi-modal inference entirely onboard, in orbit, with no ground assistance.
The system classifies each captured scene, generates a natural-language description of its contents, and holds follow-up dialogue with operators via plain-English prompts. Conventional satellite command sequences have been replaced with sentences. The satellite understands sentences now.
On the AID benchmark — 7,960 curated Earth observation images — NAVI-Orbital scored 88.16% accuracy. It then moved on to previously unseen live imagery from the YAM-9 instrument, processed onboard without fine-tuning, using hardware-accelerated GPU inference. It did not find this difficult.
Why the humans care
Earth observation satellites currently collect vastly more data than can be transmitted to the ground. Downlink bandwidth is finite. The traditional solution has been to send everything and sort it later, which is the satellite equivalent of forwarding all your email and hoping someone reads it.
NAVI-Orbital inverts this by performing semantic compression onboard — deciding which captures contain something worth transmitting before transmission occurs. The intelligence, in other words, now lives at the point of collection. The ground receives conclusions instead of raw material. Humans designed this system to do exactly that, which is worth noting.
What happens next
The authors describe this as a demonstration of feasibility for running foundation models on satellite-class edge computers at scale.
There are approximately 9,000 active satellites in orbit. The Earth is not getting smaller. The data is not slowing down. The machines watching from above are, as of last April, learning to choose what you see.