Meta Platforms Inc. has teamed up with Booz Allen Holding Corp., a U.S. government contractor, to develop an artificial intelligence system for the International Space Station.

The companies detailed the project today. Space Llama, as the AI system is called, is based on a fine-tuned version of Meta’s Llama 3.2 language model series. It’s designed to support science projects in the ISS National Laboratory, a research lab located aboard the station.

Certain types of materials are easier to mix in low-gravity conditions. As a result, astronauts use the ISS National Laboratory to study material combinations that would be difficult to create on Earth. It’s believed that research into such compounds could enable new pharmaceutical manufacturing techniques.

The ISS National Laboratory’s work also spans a wide range of other areas. Astronauts have studied how systems such as 3D printers, robots and sensors operate in space. Additionally, the lab is used to process data from the numerous Earth observation instruments attached to the ISS.

Meta debuted Llama 3.2, the language model series that powers Space Llama, last August. It comprises four open-source models with 1-90 billion parameters. The Llama 3.2 series is optimized to run on systems with limited processing capacity, which makes it well-suited for the computing equipment aboard the ISS.

Space Llama runs a fine-tuned version of Llama 3.2 on an appliance called the Spaceborne Computer-2. The system was made by Hewlett Packard Enterprise Co. in collaboration with NASA. It combines HPE servers with Nvidia Corp. graphics cards, about 130 terabytes of flash storage and NASA-developed power management equipment.

According to Meta, Space Llama uses Nvidia’s cuDNN and cuBLAS libraries to run the onboard AI models. The former library provides prepackaged software building blocks for building neural networks. The cuBLAS toolkit, in turn, helps developers boost their AI models’ performance.

Meta says astronauts can use Space Llama to extract data from technical reference documents. Additionally, the system’s Llama 3.2 implementation lends itself to processing multimodal files thanks to built-in computing vision features. The AI doesn’t require instructions from Earth to work, which means it’s capable of performing calculations even when network bandwidth is limited.

“This initiative is set to support the ISS National Laboratory’s researchers in a wide range of scientific projects and represents a critical step for lunar and Mars exploration, enablement of modern satellite and drone capabilities, and the next generation of autonomous systems,” Meta staffers wrote in a blog post today.

Image: Meta