Welcome to Issue 21, where ESA opens a public Earth model as two onboard tests report from orbit
No orbital-compute story cleared the bar this week, so this issue stays with AI applied to space.
The European Space Agency opened a public question-and-answer model for Earth observation, with 24 billion parameters. The same day, a research team reported the first in-orbit run of a marine anomaly detector on PhiSat-2. The night before, a Korean cubesat launched with a commercial edge-AI device and a module that records what that model does.
Specifics below.
Did you know?
We constantly update our Bits & Orbits Insights!
Check out the growing set of independent dashboards, trackers, and maps on the
AI × Space frontier:
AI4Space
ESA launches Earth Virtual Expert (EVE), a 24B-parameter LLM for Earth observation knowledge
On October 8, 2026, the European Space Agency released Earth Virtual Expert (EVE), a question-and-answer platform built to make Earth observation and Earth science knowledge accessible to researchers, policymakers, journalists and the public. The core model, EVE-Instruct, is a 24-billion-parameter LLM built on Mistral's Small 3.2 and further optimised for reasoning and question answering.
EVE was funded by ESA Φ-lab and built by Pi School, in partnership with the French AI company Mistral AI and Imperative Space, a firm working in innovation, communications and education for the space sector. The team curated a domain corpus by manually selecting 172 sources across 22 trusted publishing institutions, spanning open-access, private and proprietary collections, the last via a partnership agreement with Wiley. In benchmarking, EVE-Instruct was compared against its parent model and three LLMs of comparable scale, Gemma3, Qwen3 and Llama4 Scout, and led on the multiple-choice question-answering tests and on open-ended QA.
To curb hallucination, EVE runs a retrieval-augmented generation pipeline that grounds answers in curated knowledge bases. EVE-Instruct first acts as an evaluator, producing a binary hallucination label with a justification. If a hallucination is flagged, the query is reformulated, new evidence is retrieved, and the model ranks the original and revised answers to return the better-supported one.
A six-month pilot began in September 2025 with 350 users working through a graphical interface and an API. The model weights, code, curated corpus, evaluation benchmarks and a subset of the synthetic fine-tuning data are now published on Hugging Face and GitHub under ESA Φ-lab's FutureEO programme. For now EVE is text-only and does not reason over imagery or geospatial data directly; the team plans to extend it into a multimodal, agentic platform and has built it on the Model Context Protocol to connect external geospatial tools.
We covered Ameca and Frankie as conversational front-ends for EVE in Issue 19; this is the public launch that demo said was still ahead.
First in-orbit demo of onboard marine anomaly detection on ESA's Φsat-2
Φsat-2 mission patch. Credit: ESA
Also on October 8, 2026, a team led by Clotilde Szywala at IRT Saint Exupéry posted a preprint (arXiv:2610.11735) reporting the first in-orbit demonstration of an onboard marine anomaly-detection pipeline on the European Space Agency's Φsat-2 CubeSat. A companion preprint (arXiv:2610.03649) posted October 2 by Thomas Goudemant and co-authors, including Szywala, Benjamin Francesconi, Michelle Aubrun, Yves Bobichon, Marjorie Bellizzi, and Adrien Girard, details the pipeline design behind the demonstration.
The pipeline runs in stages on the satellite. It first segments sea from land, then uses a self-supervised encoder to compress marine regions into a latent representation, and flags anomalies as deviations from a learned normal sea state, with optional characterization of specific event types. The approach targets oil spills, harmful algal blooms, and sediment plumes, and reduces downlink by transmitting only anomalous tiles rather than full scenes.
Φsat-2 is a 6U CubeSat in a sun-synchronous orbit at roughly 500 km, carrying a sensor with 5-meter resolution across 8 visible and near-infrared bands, with inference running on a Ubotica CogniSat processor built around an Intel Movidius Myriad 2 accelerator. The same pipeline was developed for deployment across multiple missions, including the Microsoft/Thales Alenia Space IMAGIN-e "space edge computing" demonstration on the International Space Station.
Before launch, the pipeline was trained and validated on simulated Φsat-2 imagery to check both algorithmic performance and compatibility with the resource-constrained onboard hardware. After integration and functional validation in orbit, early experiments on real Φsat-2 acquisitions revealed a significant mismatch between the simulated and in-orbit data. The authors conclude that simulation-based development is useful for pre-flight risk reduction, but that reliable scientific validation requires representative in-orbit data and robust annotation and sensor-aware design, and must be distinguished from functional validation.
ArbaLabs flies an AI-execution checker beside a commercial edge-AI device on SD-1Sat
On October 7, 2026, South Korea's Nuri rocket flew for the fifth time from the Naro Space Center and placed 15 satellites in orbit. One of them is SD-1Sat, a 3U cubesat from Space & Bean. It weighs 4.1 kilograms and is built for a sun-synchronous orbit lasting six months to a year.
Space & Bean's own mission is the enclosure. SD-1Sat tests whether a protective housing can shield commercial off-the-shelf electronics from radiation, heat, and vibration, so operators can skip space-rated parts. Inside that housing sit a commercial edge-AI device, an optical camera, and Internet-of-Things communications hardware. Radiation sensors behind two shielding materials will compare dose, with attention on the South Atlantic Anomaly. The satellite was built with Yonsei University, DRB, and ArbaLabs.
The AI piece is ArbaLabs' ArbaEdge, also called AE-PX1. It does not run the satellite's model. It sits beside the on-device inference stack, records which model ran, and stores a fingerprint of the output so someone can check the result later without a cloud connection. ArbaLabs says the demonstration is underway and has not published results. No source names the model on the commercial edge-AI device.
Ashley Reeves, ArbaLabs' chief executive, told Payload that few people are putting checks on what these models do, and that space is a beachhead for the same verification product on autonomous vehicles, drones, and humanoid robots.
Till next time,
Meta-beat Column of this week
Read also about the AI Pipeline that sits at the core, producing this Newsletter, including its ups and downs of this week: