NASA Flies Google's Gemma 3 Model in Orbit to Read Satellite Images
NASA and Loft Orbital flew a compressed Gemma 3 model on the YAM-9 satellite after it hit 88 percent accuracy in ground tests, then read live images in orbit.
Why NASA put a language model in orbit
NASA's Jet Propulsion Laboratory and the satellite operator Loft Orbital wanted to know whether a general-purpose vision-language model could describe what a satellite sees without needing a dedicated, retrained model for every new target. Their answer came from a project called NAVI-Orbital, an agentic software framework built by JPL's Juan M. Delfa, Taran Cyriac John, and Loft Orbital's Andrew W. Herson. The team deployed a 4-bit, compressed version of Google's open-weights Gemma 3 4B model directly onto a satellite's onboard computer, then asked it to describe what its camera saw.
Delfa framed the result as a change in how researchers can talk to a spacecraft. "Now, a scientist can write a prompt, upload it to the spacecraft, and that will be taken into account by the system," he said, contrasting it with older approaches that required an operations team to write structured commands, via IEEE Spectrum.
Inside NAVI-Orbital
NAVI-Orbital coordinates its work with a LangGraph-based conductor and keeps Gemma 3 walled off from the satellite's flight software; the model only reads images and writes plain-text descriptions, with no access to spacecraft controls. That separation let the team test the concept without touching how YAM-9 actually flies.
The system ran on an Nvidia Jetson Orin AGX, a small compute module common in robotics work. Because the model was compressed to a 4-bit, 4-billion-parameter format, it needed only 8 gigabytes of memory, letting it run on that low-power hardware. Delfa said that lightweight footprint was the point: "You can run it in a tiny, tiny computer."
What the benchmark showed
Before launch, the team validated NAVI-Orbital on the ground using a benchmark dataset of 7,960 images, where Gemma 3 reached 88 percent classification accuracy without being trained or fine-tuned on that dataset's specific categories, the same base Gemma 3 model anyone can download and run on a laptop. Once NAVI-Orbital was in orbit, NASA ran two live capture tests with a camera on YAM-9: one over Toulouse, France, and one over the coast of Argentina. Gemma 3 produced a text description of each image and answered a set of scripted questions, such as whether an image showed commercial or residential areas.
YAM-9 itself carries several radiation-hardened processors and runs on solar panels that supply between 150 and 500 watts, depending on the satellite's position.
Why this matters for bandwidth
Loft Orbital general manager Paul Lasserre described the appeal as "semantic compression": rather than downlinking full images, a satellite can send a short text summary instead. "It doesn't matter if the link is slow, because you're downlinking dozens of kilobytes instead of dozens or hundreds of megabytes," Lasserre said.
Delfa pointed to wildfire detection as a real-world case. Bandwidth and processing limits can currently delay wildfire results by up to 90 minutes; onboard image analysis could remove much of that lag. Retooling NAVI-Orbital for a new target, like wildfires instead of land use, is mostly a matter of editing a text prompt rather than retraining and redeploying a new model.
Key facts
- NAVI-Orbital ran Google's Gemma 3 4B model, compressed to a 4-bit format, aboard Loft Orbital's YAM-9 satellite.
- On a ground-tested benchmark of 7,960 images, the system reached 88 percent classification accuracy with no extra fine-tuning.
- The compressed, 4-billion-parameter model needed only 8 gigabytes of memory and ran on an Nvidia Jetson Orin AGX module.
- YAM-9 draws 150 to 500 watts from its solar panels, and onboard analysis could cut wildfire-detection delays of up to 90 minutes.
Check your understanding
- What accuracy did NAVI-Orbital reach on the ground-tested benchmark of 7,960 images before launch?
- Why does analyzing images onboard YAM-9 with Gemma 3 reduce how much data NASA needs to send to the ground?