TerraMind
TerraMind is the first any-to-any generative multimodal Earth observation foundation model, built by IBM and ESA’s Φ-lab and launched in April 2025. It is open-sourced on Hugging Face, and lighter edge versions released in October 2025 run on laptops or onboard satellites.
TerraMindInside TerraMind
TerraMind, co-developed by IBM Research and ESA Φ-lab, is a dual-scale, any-to-any generative foundation model pretrained on hundreds of billions of tokens from the multimodal TerraMesh dataset.
It introduces a Thinking-in-Modalities approach to handle missing modalities and reports state-of-the-art results on Earth observation benchmarks such as PANGAEA.
All sizes are open-sourced on Hugging Face, and the October 2025 TerraMind.tiny and TerraMind.small variants are engineered to run on normal or edge hardware, including laptops and onboard satellites, enabling on-device inference.
Recent developments
- Oct 2025TerraMind.tiny and TerraMind.small edge versions released for laptops and onboard-satellite inference.
- Apr 2025IBM and ESA Φ-lab launched TerraMind, an open any-to-any multimodal EO foundation model.
Explore TerraMind for a project
Check the provider’s current capabilities, access terms and documentation before selecting a tool or planning an integration.