NASA Runs Google's Gemma 3 Aboard Orbiting Satellite for On-Board Image Analysis

NASA's Jet Propulsion Laboratory has run Google's Gemma 3 large language model in orbit, achieving what the agency describes as the first in-orbit demonstration of a vision-language model analyzing imagery captured by a satellite's own sensor.

The system, called NAVI-Orbital, deployed a 4-bit quantized version of Gemma 3 4B — Google's open-weights model, the same one available for download on Hugging Face — aboard a YAM-9 satellite built by Loft Orbital. No fine-tuning or additional training was performed on the model before or during the mission.

How it works

NAVI-Orbital is an agentic software framework built by NASA JPL's Juan M. Delfa and Taran Cyriac John alongside Loft Orbital's Andrew W. Herson. It uses a LangGraph-based conductor to coordinate operations and runs Gemma 3 to produce plain-text descriptions of satellite imagery.

On the ground, before launch, the system was benchmarked against a dataset of 7,960 images and reached 88 percent classification accuracy, according to the research team's paper. In orbit, NASA conducted two live capture tests — one over Toulouse, France, and another over the coast of Argentina — prompting Gemma 3 with scripted questions about whether images showed commercial or residential areas, or natural features.

The model ran on an Nvidia Jetson Orin AGX, a low-power compute module common in robotics applications. The 4-bit, 4-billion-parameter model requires only 8GB of memory, according to Delfa, which is what made it feasible to run on the satellite's hardware. YAM-9 itself carries a compute cluster of radiation-hardened FPGAs, CPUs, and GPUs shared across multiple customer payloads, powered by solar panels generating between 150 and 500 watts depending on orbital position.

Why bandwidth matters

Paul Lasserre, general manager at Loft Orbital, called the approach a form of "semantic compression." Rather than downlinking full-resolution images — which can run into the hundreds of megabytes — a satellite using this approach can transmit a text summary instead, at a fraction of the data size.

"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 practical example. Current satellite-based wildfire detection can face delays of up to 90 minutes due to downlink bandwidth and ground-based processing constraints, according to research cited by the team. An on-board model capable of flagging a fire directly could, in theory, cut that delay significantly — though NASA has not yet demonstrated this specific use case in orbit.

Not yet flying the spacecraft

NAVI-Orbital is deliberately isolated from flight software; it reads images and generates descriptions but has no access to spacecraft controls. Delfa frames the demonstration as a shift in how researchers interact with orbiting hardware. Retargeting the system for a new task — say, searching for wildfires instead of classifying land use — requires only editing a text prompt rather than rewriting, retraining, and revalidating onboard software, according to Delfa.

Delfa also said the NAVI framework was conceived partly as groundwork for AI companions that could assist astronauts operating under the physical constraints of a spacesuit, though this application remains conceptual.

What this means

This is a narrow but concrete proof of concept, not a leap toward autonomous spacecraft. The notable technical fact is that an unmodified, publicly downloadable 4-billion-parameter model — quantized to 4-bit precision and run on commodity edge hardware — performed useful classification work in orbit without any space-specific retraining. That lowers the bar for future orbital AI deployments considerably, since mission teams won't necessarily need custom-trained models for every new sensor or target type.

The bandwidth argument is the more immediately practical one: most satellites are downlink-constrained, and swapping megabyte-scale images for kilobyte-scale text descriptions is a real operational advantage regardless of how sophisticated the underlying model is. Whether 88 percent accuracy is good enough for operational use — wildfire detection, disaster response, or defense applications — will depend heavily on the cost of false negatives in each use case, something this demonstration doesn't resolve.