NASA has run a large language model directly in orbit for the first time. As IEEE Spectrum reports (23 July 2026), its Jet Propulsion Laboratory loaded Google's Gemma 3 onto the YAM-9 satellite to analyze images of Earth in real time — instead of downlinking raw imagery to the ground. It is the first demonstration of a vision-language model running in space.
Key takeaways
- Model: Google Gemma 3 4B (open-weights), compressed to 4-bit, 8 GB memory, unmodified — straight from Hugging Face.
- Hardware: YAM-9 satellite by Loft Orbital, an Nvidia Jetson Orin AGX compute module, powered by 150–500 W from solar panels.
- Task: describing Earth images and answering questions about terrain (commercial and residential areas, natural features).
- Accuracy: 88% on a benchmark of 7960 images; two live tests — Toulouse, France and the Argentine coast.
- Payoff: downlinking dozens of kilobytes of text instead of hundreds of megabytes of imagery.
What flew to orbit
NASA's Jet Propulsion Laboratory ran the deployment. Aboard the YAM-9 satellite, built by Loft Orbital, it launched Google's Gemma 3 4B — the open, four-billion-parameter version running on an Nvidia Jetson Orin AGX module with radiation-hardened chips. Solar panels supply the whole system with 150 to 500 watts of power.
The model fits into 8 GB of memory thanks to 4-bit compression. Its job is to analyze images of Earth: generating text descriptions and answering questions about what is in frame — from commercial and residential areas to natural features. Two live captures were run: over Toulouse, France, and off the Argentine coast.
Semantic compression instead of raw images
The key benefit is not the image recognition itself, but what reaches the ground. Instead of pulling down hundreds of megabytes of raw images, the satellite sends a concise text description of what it saw. That changes the economics of the radio link, which is the bottleneck in Earth observation.
It doesn't matter if the link is slow, because you're downlinking dozens of kilobytes instead of dozens or hundreds of megabytes.
Paul Lasserre, general manager at Loft Orbital.
A model straight from Hugging Face
What is surprising is how little had to change. The team used an unmodified version of Gemma 3 downloaded from Hugging Face — with no space-specific fine-tuning. On a benchmark set of 7960 images the model reached 88% accuracy. By comparison, classic Earth observation requires downlinking all the data to the ground and processing it after the fact, which lengthens response time and loads the link.
The second change is programmability. The satellite can be re-tasked with a single text prompt, without uploading new software.
This is a major shift. Now, a scientist can write a prompt, upload it to the spacecraft, and that will be taken into account by the system.
Juan M. Delfa, technical group lead at NASA JPL.
Why it matters
The experiment pushes computation to the very edge — literally into orbit. The fact that a small, open, consumer-grade model sufficed, rather than a costly model built specifically for space, shows how commoditized capable models have become. Not long ago such a payload would have meant years of work on a bespoke system.
The biggest value, though, is flipping the economics of bandwidth. In Earth observation the constraint is not compute but the radio link — processing in orbit and sending only the conclusions sidesteps it. Add prompt programmability, which turns the satellite from a rigid sensor into a device you can re-task in flight. Still, this is a demonstration: a 4B model and 88% accuracy are a showcase ceiling, bounded by power and memory budgets.
What's next
- Technical details are described in the NAVI-Orbital preprint on arXiv — the basis for replication attempts by others.
- A 4B model and 88% accuracy are a demonstration ceiling — larger models are limited by the power budget (150–500 W) and 8 GB of memory.
- Loft Orbital positions YAM-9 as a platform for AI in orbit — future payloads depend on operators' decisions.
