Meta's open-source AI is blasting off to the ISS

Space Llama is a custom offline AI system built by Meta and Booz Allen for the International Space Station (ISS). Running on a fine-tuned version of Meta's open-source Llama 3.2, it is engineered to handle multimodal intelligence and research tasks without relying on an internet link back to Earth.

On the ISS, reliable internet access is essentially nonexistent, and transmitting heavy files back to mission control introduces painful bandwidth caps and lag. Space Llama solves this by packing frontier AI directly onto local space-grade hardware, keeping sensitive data strictly on board while giving astronauts instant access to technical intelligence.

The hardware stack in orbit

Deploying a multimodal model inside a cramped, power-constrained space laboratory takes serious engineering.

The deployment brings together Booz Allen’s edge software A2E2 (AI for Edge Environments), Hewlett Packard Enterprise’s Spaceborne Computer-2, and Nvidia accelerated computing powered by CUDA, cuDNN, and cuBLAS, combined with Meta’s Llama vision features.

Using a low-power, compact architecture similar to satellite computing boxes, the team managed to slash processing times from several minutes down to just over a single second. It marks the first time a multimodal stack of this caliber has operated beyond Earth's atmosphere.

Offline search and system repairs

Space Llama is built to digest multimodal inputs, spanning text, visual cues, and audio.

Its most immediate everyday use case is helping astronauts navigate thousands of pages of dense flight rules, equipment repair manuals, and scientific protocols. Instead of paging through reference docs or waiting on a radio link with Houston, researchers can simply query the model locally to surface maintenance steps.

Booz Allen CTO Bill Vass highlighted that space exploration has long been throttled by reliance on ground-based communications. Pushing compute directly into orbit unlocks rapid troubleshooting and agile data access right at the operational edge.

The initiative builds on preliminary tests from August 2024, when Booz Allen first operated an offline large language model on HPE’s Spaceborne Computer-2 aboard the station. Meta subsequently opened Llama up to US government and defense contractors in November 2024, clearing the runway for Space Llama.

How to run similar edge setups

You cannot download the exact Space Llama orbital build from a public repo, as the fine-tuned weights and edge wrappers remain restricted to station operations.

However, because the underlying engine is Meta's open-source Llama 3.2, developers can replicate the exact same offline architecture on Earth. The base model weights are openly accessible on Hugging Face and Meta's developer portal, ready to run on local edge machines using tools like Ollama or vLLM.

The main friction point for terrestrial builders is hardware sizing: packing multimodal Llama onto small local chips requires careful quantization and memory tuning, though thankfully Earth-bound machines do not have to endure cosmic radiation.

Why it matters

Space Llama is the ultimate stress test for edge AI, and it proves that open-source models hold a massive tactical advantage over cloud-tethered alternatives.

When you are operating in deep mines, offshore vessels, remote defense outposts, or low Earth orbit, closed APIs that depend on continuous fiber connections simply do not work. You need open weights that can be inspected, hardened, and run on local silicon without sending telemetry back to base.

Meta gets to showcase that Llama is not just a consumer toy, but mission-critical infrastructure capable of powering deep-space missions to the Moon and Mars. It draws a clear competitive boundary: where the internet ends, open-source AI begins.