Slide 1
A twenty-one-million-parameter open model, built by an academic team at Cambridge, just out-mapped Google DeepMind's flagship Earth model. It's called TESSERA v2, and how they got there is stranger than the result.
One learned vector for every 10-meter pixel of the planet
TESSERA's unit isn't a satellite tile. It's a single ten-meter pixel. It takes a full year of radar and optical observations and squeezes them into one frozen vector you can probe for any map.
Pretraining loss barely predicts the skill
Across three hundred and ninety-five controlled runs on a thousand GH200 chips, pretraining loss barely predicted downstream skill. The correlation was negative zero point one eight. Chasing a lower loss can cost about two and a half times more compute for the same result, so they built the scaling law around downstream performance instead.
Best composite of any system on the 29-task suite
On the full twenty-nine-task benchmark, the twenty-one-million-parameter model scores zero point six one one. The best composite of any system they tested, ahead of TESSERA v1, AlphaEarth, and OlmoEarth. And on AlphaEarth's own fifteen tasks, it still wins, zero point five eight one to zero point five six.
Sixteen numbers keep about 92% of the score
How does a model this small compete? They distill a one-billion-parameter teacher into nested embeddings. The first sixteen numbers alone keep about ninety-two percent of the full score, at one-eighth the storage. The naive version fails; distilling against a fixed teacher supplies the ordering.
What the number doesn't tell you
Three honest caveats. The v2 weights aren't out yet. The team says v2 lands soon, but today you can only grab version one. The scores come from the paper's own protocol, not a re-run of Google's internal tests. And this is satellite radar and optical only. Nothing here transfers to general vision.
Read it now, run v2 soon
The paper's on arXiv today, and the code is MIT-licensed. Version one's embeddings are already free through GeoTessera, and the v2 weights land at the same repo when they're ready.
Slide 8
The lesson underneath the maps: optimize for the task you actually care about, not the number that's easy to watch. That's how twenty-one million parameters caught a Google flagship.
