On‑satellite vision‑language fine‑tuning
Engineers are fine‑tuning compact vision‑language models (e.g., LFM2.5‑VL‑450) on satellite datasets like VRSBench to enable lightweight on‑satellite object detection and QA without full image downloads. The approach aims to emit small JSON detections from orbit rather than streaming raw imagery to ground. (x.com)
Satellites usually send pictures to Earth first and analyze them later; engineers are now training small vision-language models to do more of that work in orbit. (docs.liquid.ai) Liquid AI published a new example in April 2026 for fine-tuning its 450 million-parameter LFM2.5-VL-450M model on satellite imagery tasks using VRSBench, a remote-sensing benchmark accepted to the NeurIPS 2024 Dataset and Benchmark Track. (liquid.ai) (arxiv.org) VRSBench packages three jobs that matter for Earth-observation images: 123,221 visual question-answer pairs, 52,472 object references with boxes, and 29,614 detailed captions across 29,614 images. (arxiv.org) A vision-language model is a system that reads pictures and text together, like a captioning tool that can also answer questions and point to objects; Liquid says LFM2.5-VL-450M adds bounding-box prediction and function-calling support. (liquid.ai) (huggingface.co) That combination lets a satellite turn an image into a small structured output, such as a list of detections, instead of sending every full-resolution frame to the ground. Liquid’s model card says the model is built for edge deployment, supports image tiling, and can be exported in formats including ONNX and quantized variants for local inference. (liquid.ai) (huggingface.co) The bottleneck is not new. A 2024 NASA briefing said Earth-science missions need onboard data reduction because sensor bandwidth can exceed downlink bandwidth, and NASA said in 2025 that onboard algorithms can prioritize what gets transmitted to Earth. (nasa.gov) (esto.nasa.gov) Space agencies are already testing narrower versions of the same idea. The European Space Agency’s PhiSat-2 cubesat, launched on August 16, 2024, runs onboard artificial-intelligence applications that can discard cloud-obscured images and detect maritime vessels before transmission. (earth.esa.int) (esa.int 1) (esa.int 2) Liquid’s example is aimed at developers rather than flight-qualified spacecraft teams. Its tutorial runs data preparation in a cloud container, downloads about 12 gigabytes of VRSBench data, and launches fine-tuning on an Nvidia H100 graphics processor, with checkpoints saved to cloud storage. (docs.liquid.ai) That leaves a gap between a working demo and a satellite deployment. NASA’s 2024 survey of space artificial intelligence said onboard models must also contend with radiation-induced bit upsets and other reliability constraints that do not apply to ordinary edge devices. (nasa.gov) The immediate shift is simpler: move from “download the picture, then ask questions” to “ask the question on the spacecraft, then downlink the answer.” For operators paying for bandwidth and time, a compact model that emits detections instead of raw imagery points at that trade. (docs.liquid.ai) (nasa.gov)
