FUNDING · MODELS · #2427
OlmoEarth v1.3 — open satellite embedding model and global 10m embeddings
OlmoEarth v1.3 is released as an open satellite embedding model that produces 128‑dimensional (and a 64‑dim subset) embeddings at 10‑meter resolution for every pixel on Earth, with precomputed annual global embeddings, open weights/code/data, and a browser-based OlmoEarth Embeddings Explorer. The team reports v1.3’s 128‑dim embeddings lead on 20 of 27 evaluations (64‑dim leads on 18 of 27), introduces the PLANTEUR crop benchmark, and says the model was cheaper to train (final run ≈ $2,720 of rented H100 time) with lower compute and carbon than prior versions.
KEY POINTS
- OlmoEarth v1.3 is released as an open satellite embedding model that produces 128‑dimensional (and a 64‑dim subset) embeddings at 10‑meter resolution for every pixel on Earth, with precomputed annual global embeddings, open weights/code/data, and a browser-based OlmoEarth Embeddings Explorer.
- The team reports v1.3’s 128‑dim embeddings lead on 20 of 27 evaluations (64‑dim leads on 18 of 27), introduces the PLANTEUR crop benchmark, and says the model was cheaper to train (final run ≈ $2,720 of rented H100 time) with lower compute and carbon than prior versions.
- Open, high‑quality global embeddings let practitioners and researchers build lightweight, low‑cost satellite mapping and change‑detection workflows without heavy compute or reprocessing raw imagery.
WHY IT MATTERS
Open, high‑quality global embeddings let practitioners and researchers build lightweight, low‑cost satellite mapping and change‑detection workflows without heavy compute or reprocessing raw imagery.