A language model that reads the Moon.

Four maps, one tile.

The model never sees a photograph alone. Every tile arrives as a stack of co-registered rasters covering the same 51 × 67 km patch of surface — what it looks like, what shape it is, what the maps already say about it, and what its spectra imply about composition and maturity. Switch between them below; the terrain underneath does not change, only the way of looking at it.

Mare · Im2 · 47.9°N 25.9°W WAC reflectance

Terrain

Mare Highland

Raster

WAC Elevation Geologic map Clementine ratio

WAC reflectance

LROC Wide Angle Camera imagery — the surface as it appears to the eye. Albedo separates dark mare basalt from bright highland anorthosite, and picks out fresh ejecta.

Ask the model.

Pick one of eight held-out tiles and talk to the deployed model. Every response is a real generation, and every number in it is checked against the rasters the model was shown.

Each message runs a real generation on a single GPU that scales to zero between visitors, so the model has no memory of previous turns — every answer is produced from the tile and your instruction alone. The fact check under each response is computed from the rasters, not by the model.

Under the hood.

The architecture is deliberately standard — a frozen vision tower, a small trainable bridge, an adapted language model. Nothing here is a new block. Walk the six stages a tile passes through on its way to a sentence.

Coordinates are deliberately withheld from the prompt. Given exact latitude and longitude the model lowers its loss through a location-to-geology shortcut, and the vision gate never opens — it stops looking at the tile.

What it's trained to say.

If the supervision were prose alone, the loss could bottom out on generic connective tissue — “I view this landscape as…” — which a 7B language model can write without ever looking at the tile. So the targets carry deterministic, tile-specific facts, and the tokens that spell them out carry six times the weight in the cross-entropy.

Below is a real response, run through the model's own token-weighting patterns. Toggle a category to see which words survive — those are the ones that cannot be produced without reading the rasters.

The highlighting runs the same three patterns the training code uses to build its token weights: named units and formations, model ages in Ga, and image-only measurements — albedo, relief and roughness in metres, the 415/750 ratio, crater densities. Getting this tile's numbers right becomes the cheapest way to lower the loss, which is what forces the model to use the bridge instead of collapsing onto generic prose.

What the measurements say.

Grounding is treated as a property to be measured, not asserted — which means the same protocol that demonstrates the model works also shows precisely where it stops working. Three measurements, in the order they should be read.

The conclusion this points to. Target design governs what a grounded planetary vision-language model states and where it stays silent. Numeric crater-chronology ages belong in decoupled heads reading the same representations, or in retrieval where published records exist — not in generated text. The failure sits in the language extraction pathway, not in the representation, which is a considerably more fixable place for it to be.

Paper, code & weights

Data credits: LROC WAC imagery and LOLA topography (NASA/GSFC/ASU), the Unified Geologic Map of the Moon (USGS SIM 3116), Clementine UVVIS (NASA), and published mare-basalt model ages after Hiesinger et al. Stated crater-count ages on mare tiles reflect regional published chronology rather than independent tile-level dating; highland tiles are correctly left undated.