First vision-language model deployed on an orbital satellite

In April, an Earth observation satellite autonomously identified its target without human analysts on the ground for the first time. The milestone marks the first reported use of a vision-language model (VLM)—an AI model that combines text and image analysis—in orbit. Space infrastructure company Loft Orbital achieved the demonstration on its YAM-9 spacecraft using Google DeepMind’s Gemma 3 vision-language model.

The demonstration relied on NAVI-Orbital, a software package built by the NASA Jet Propulsion Laboratory (JPL) to run the model on-orbit. Typically, satellites transmit large volumes of raw data to Earth for ground-based analysis. Running a VLM directly on the spacecraft allows for on-board data triage. Google DeepMind’s Gemma 3 model is designed for edge applications, meaning it runs on limited hardware far from a data center. Paul Lasserre, Loft’s head of AI, stated that the use of a VLM opens the door to always-on, patrol layers in space. Lasserre noted that with a VLM, operators can use natural language logic to instruct satellites to monitor specific areas and flag suspicious activity.

Loft operates as an infrastructure-as-a-service provider rather than a traditional satellite manufacturer. The company currently operates 12 spacecraft on orbit, which includes YAM-9, launched in the fall of 2025. Loft also recently built six new satellites for EarthDaily, a customer that will analyze and market the collected data.

To achieve real-time coverage of anywhere on Earth, Lasserre estimates it would take somewhere between 50 and 100 satellites like YAM-9. “Now that we’ve proven the concept, that’s really the direction of travel,” said Paul Lasserre, Loft’s head of AI. Other space companies are also pursuing orbital AI. Planet Labs is researching AI applications, including VLMs, for its satellites. Meanwhile, Kepler Communications, which operates GPUs in space, launched spacecraft in January and has noted undisclosed compute use cases, though it declined to confirm VLM deployment due to non-disclosure agreements.

The software development also has implications for deep space exploration. Juan Delfa Victoria, a technical leader in NASA JPL’s AI group, and JPL researcher Taran Cyriac John originally conceived the underlying software concept as an interactive assistant for astronauts exploring the moon or Mars. Delfa Victoria described the goal as providing an interactive AI assistant, similar to those in video games and movies, to help astronauts who cannot easily use keyboards while wearing pressurized suits.

Why it matters

Running AI models directly on satellites could make space-based sensors far more useful by performing initial data triage in orbit, reducing the flood of raw data that analysts currently have to wade through. In the longer term, the demonstration serves as a proof point for deploying larger-scale AI infrastructure in space.