TESSERA v2 shows pretraining loss barely predicts EO downstream

A Cambridge-led group just published what they call the largest controlled scaling study for Earth-observation foundation models to date, and the punchline undercuts a heuristic most teams in the space still lean on. In the TESSERA v2 preprint on arXiv, the authors report 395 training runs on 1,024 GH200 superchips within a fixed pixel-wise Barlow Twins family, each evaluated on 15 downstream tasks, and find that pretraining loss barely predicts downstream performance, with |Pearson r| under 0.2.

The practical read is blunt: selecting checkpoints by loss, which is what you do by default when you spin through a sweep, wastes a large share of the compute. The authors also offer a compute rule that fell out of the sweep, that as the training budget grows, the encoder and the data should grow together while the projector stays fixed. That is the kind of simple allocation rule that mid-sized EO teams can actually adopt without redoing their own scaling law.

The other headline is the distillation result. The team trains 0.5B and 1B pixel-wise models, with a 2B still in training, then distills them into a 21-million-parameter student they call TESSERA v2-1B-M. The claim is that in aggregate this student outperforms all open and proprietary models they tested, some of which are orders of magnitude larger. The students use Matryoshka representations, so a 16-dimensional prefix keeps 92% of the full 128-dimensional performance at 1/8 the storage, which matters a lot when you are serving embeddings as data across a planet.

The honest caveat is that everything above sits inside one family of pixel-wise Barlow Twins models on one set of 15 downstream tasks, so how far the loss-vs-downstream decoupling generalises to other EO pretraining paradigms is not something the paper actually settles. What the reporting also does not give you is how the 21M student behaves outside those benchmarks or how expensive the distillation step is relative to pretraining the teachers.

Still, if the finding survives replication, the near-term winners are climate, agriculture, and infrastructure monitoring pipelines that have been eyeing proprietary EO foundation models, since a 21M-parameter student they can host themselves is a very different procurement conversation. Cambridge is also promising to release global TESSERA v2 embeddings covering 2017-2025 once training finishes, which would let downstream teams skip pretraining entirely.