TerraMind
IBMTerraMind is an any-to-any generative Earth-observation model developed by IBM, ESA Phi-Lab, and Forschungszentrum Juelich. It processes six satellite and terrain modalities, learns both pixel- and token-scale representations, and can generate missing modalities or supply embeddings for fine-tuned downstream models through TerraTorch.
Verified 2026-09-10 via the Hugging Face model card and IBM repository.
Openness
3 high confidence- license
- Apache-2.0(OSI)
- weights
- open(Apache-2.0 ungated checkpoint)
- data
- open(TerraMesh dataset published on Hugging Face)
- code
- partial(inference, generation, and fine-tuning examples without the full pretraining pipeline)
TerraMind publishes its weights and named training dataset, but the product repository documents use and adaptation rather than an end-to-end pretraining run.
- https://huggingface.co/ibm-esa-geospatial/TerraMind-1.0-large/raw/main/README.md recorded 2026-09-10
Apache-2.0 card linking TerraMesh, describing 500B tokens from 9M aligned multimodal samples, and documenting generation and TerraTorch fine-tuning.
Adoption
1 high confidence2,922 downloads in the trailing 30 days for the declared large checkpoint.
- https://huggingface.co/api/models/ibm-esa-geospatial/TerraMind-1.0-large recorded 2026-09-10
downloads: 2922; gated: false; private: false.
Capability
5 high confidenceTerraMind is the top anchor because one primary card directly establishes all four attributes without relying on downstream task-count prose.
- https://huggingface.co/ibm-esa-geospatial/TerraMind-1.0-large/raw/main/README.md recorded 2026-09-10
Documents config-driven fine-tuning, six raw modalities, any-to-any generation, and a dual-scale encoder-decoder operating at pixel and token levels.
- https://raw.githubusercontent.com/IBM/terramind/main/README.md recorded 2026-09-10
Provides fine-tuning configurations, multimodal generation notebooks, and tokenizer reconstruction examples for the released family.
Verified 2026-09-10