TL;DR — NASA's JPL flew a 4-bit quantized, unmodified Gemma 3 4B on a Loft Orbital satellite this year, running vision-language inference on live Earth imagery with no ground-based datacenter in the loop. This episode looks at why open weights, not APIs, are what makes onboard AI in orbit possible, and where satellite-class edge compute is already running smaller open models today.
Every model you've read about in this series so far assumes you can, in principle, ping a server. Space breaks that assumption on purpose. A satellite in Low Earth Orbit has a ground station window measured in minutes, a power budget measured in watts, and a round trip to mission control that can take hours. If your AI needs a datacenter, it doesn't fly. This is the one environment where "open-weight and runs locally" stops being a preference and becomes the only option on the table.
The proof point: Gemma 3 in orbit
On April 16, 2026, NASA's Jet Propulsion Laboratory ran what appears to be the first in-orbit demonstration of a vision-language model doing autonomous multi-modal inference entirely onboard a spacecraft. The system, called NAVI-Orbital, deployed a 4-bit quantized version of Google's open-weight Gemma 3 4B — the same checkpoint anyone can pull from Hugging Face — aboard a Loft Orbital YAM-9 satellite, with no fine-tuning performed before or during the mission, according to the arXiv paper describing the demonstration.
The numbers are the story here. The model hit 88.16% accuracy on a curated 7,960-image version of the AID aerial imagery benchmark. It ran on an Nvidia Jetson Orin AGX drawing power from a satellite bus generating 150 to 500 watts depending on orbital position, and the 4-bit quantized model needed only about 8GB of memory — small enough to fit that edge hardware without modification, as reported by TPS's coverage of the mission. Detection outputs — a predicted label, a free-text scene description, and metadata — occupy 700 to 1,060 bytes in JSON. The corresponding full-resolution captures from the YAM-9 imager run 9 to 14MB. That's roughly four orders of magnitude of compression, achieved not by shrinking the image but by replacing it with language.
TechCrunch's reporting quotes Loft Orbital's head of AI, Paul Lasserre, describing the implication plainly: "It opens the door to always-on, patrol layers in space... you can have logic — like 'monitor this border for me, and let me know when something is suspicious,' and interact back and forth with the satellites." The orchestration layer running underneath Gemma 3 is a LangGraph state machine coordinating separate detection and dialogue agents — the same open tooling you'd use to build an agent on the ground, just deployed 500 kilometers up.
It's not the only satellite running a model
NAVI-Orbital is the headline case, but it's part of a small, growing pattern rather than a one-off stunt. On the International Space Station, the IMAGIN-e payload has demonstrated onboard inference with compact variants of a Vision Transformer-based geospatial foundation model, purpose-compressed to fit resource-constrained flight hardware while preserving downstream task accuracy across five Earth-observation tasks, per a separate arXiv writeup of that ISS demonstration.
Separately, Syntiant and Novi Space demonstrated real-time, ultra-low-power AI object detection in orbit in March 2026, running quantized neural networks on an AMD Versal SoC with a dual-core ARM Cortex-A72 aboard Novi's SP240 onboard computer, according to SatNews's coverage. That mission reported the ability to retrain and hot-swap models in under 24 hours — switching a satellite's job from wildfire detection to maritime tracking without a hardware change. Different scale, different model family, same underlying principle: small, self-contained, retrainable-on-the-ground models beat anything that needs a live connection to a bigger brain.
Why open weights specifically, not just "small models"
You could, in theory, build a small proprietary model for this. What open weights buy you in orbit is control over the whole lifecycle, not just inference. A satellite is a closed system for months or years at a time; when something needs to change — a new benchmark, a new sensor calibration, a new mission objective — you can't wait on a vendor's API roadmap or eat a licensing renegotiation while the spacecraft is already flying. With an open-weight model like Gemma 3, JPL's team could quantize it to 4-bit, verify accuracy on their own curated benchmark, and load it onto flight hardware with full knowledge of exactly what's running and why it behaves the way it does. No vendor terms of service govern what a satellite is allowed to do with a locally-run, locally-owned model. That distinction matters more in orbit than almost anywhere else in this series, because there is no fallback path if a dependency changes mid-mission.
The use case: semantic compression as bandwidth strategy
The actual job being solved here isn't "AI in space" as a novelty — it's a specific, unglamorous bandwidth problem. Earth observation satellites generate more imagery than they can ever downlink. A ground station pass is short, contested across multiple payloads sharing the same bus, and every megabyte costs real money and real time. The traditional model is "acquire everything, downlink everything, let analysts sort it out on the ground" — which means most of what a satellite sees never gets looked at in any useful time frame.
NAVI-Orbital inverts that: the satellite classifies each scene onboard, writes a text description, and only flags the image itself for full downlink when the description warrants a closer look. An operator can send a plain-English prompt — "look for infrastructure near railway hubs" or "flag where natural terrain meets new development" — instead of writing a bespoke detection pipeline for each new question. That's the practical payoff of a general-purpose vision-language model over a narrow, single-task classifier: mission teams don't need a new custom-trained model every time the target changes. The trade-off is honest, too — this is a proof of concept on one satellite, one imaging instrument, and one benchmark. It's not yet a general autonomy platform, and nobody involved claims otherwise; JPL's team is explicit that this is a narrow demonstration, not a leap toward independently acting spacecraft.
Who should care about this pattern beyond aerospace: anyone building triage systems for expensive-to-move data — drone swarms, remote industrial sensors, offshore rigs, disaster-response imaging — where the constraint is bandwidth and connectivity, not compute. The lesson generalizes even if the hardware doesn't: run a small open model at the edge, downlink language instead of pixels, and only pull the expensive payload when the summary earns it.
Credits & sources
Technical details and benchmark figures on NAVI-Orbital are drawn from the arXiv paper on the in-orbit VLM demonstration. Additional reporting and quotes from Loft Orbital's Paul Lasserre come from TechCrunch's coverage of the mission. Hardware and model specifications are sourced from TPS's report on the deployment. The IMAGIN-e ISS geospatial foundation model demonstration is described in this companion arXiv paper, and the Syntiant/Novi Space low-power inference demonstration is covered by SatNews. Credit to the teams at NASA's Jet Propulsion Laboratory, Loft Orbital, Google DeepMind's Gemma team, Novi Space, and Syntiant for the actual engineering behind these missions.
Tomorrow, Day 19 comes back down to Earth for a look at Xiaomi's MiMo-V2.5 — one more open-weight model worth knowing about, no orbital hardware required.
Appendix — the field in one chart
Top comments (0)