On October 6, Google DeepMind released EmbeddingGemma 2 (EG2), an open embedding model that maps text, images, video and audio into a single shared space. Within eight days, we had tested it on nearly 800 search concepts.

The concepts came from customer queries and our own evaluation set, and we ran them on two different kinds of Earth imagery: US aerial photography (NAIP) and Sentinel-2 satellite data.

In internal tests, more than half of the zero-shot searches improved retrieval precision by 10 points or more. Examples include oil storage tanks, wind turbines, open-pit mines and deforestation.

Zero-shot means these searches used no labeled training data and no custom model. A user describes what they're looking for in plain language, and the search runs against the imagery directly. Better underlying models translate into better results for those searches.

About EmbeddingGemma 2

Embedding models get less attention than chat models, but they are among the most widely used models in AI. The most downloaded model on Hugging Face in 2026 is a sentence-embedding model, all-MiniLM-L6-v2, with 1.55 billion downloads in seven months, and models under 1 billion parameters account for 83% of all downloads on the platform. Embeddings are what make search, retrieval and recommendations work, and EG2 is the newest member of that family from Google DeepMind.

Some context on EG2:

• The first EmbeddingGemma, released in September 2025, handled text only and has been downloaded more than 20 million times.
• EG2 adds images, video and audio, all in one 768-dimensional embedding space.
• It is small: the full multimodal model needs about 567 MB of memory when quantized.
• It supports more than 100 languages.
• It is released under the Apache 2.0 license, which permits commercial use. Hugging Face already lists more than 100 community fine-tunes.

That last point is why we could move so quickly. Open weights meant we could test EG2 on our own imagery the week it came out.

How we approach new models

LGND is designed to work with multiple models rather than depend on any single one. When a new model is released, we evaluate it against real geospatial queries and add it where it performs well. We are enabling text search via EG2 and updating our LGND Geo skills so the AI agents our customers use can pick the best model for each question.

Our customers and their AI agents can already request embeddings (Clay, FarSLIP and AlphaEarth), and we’ve also proactively pre-computed embeddings for some regions so they can be searched immediately, including California's 2020 and 2022 NAIP imagery (37 million embeddings). We’ll be actively expanding coverage based on our customer needs.

What's next for LGND Geo

Here are two key advancements we're currently focused on for LGND Geo:

• Text search with EG2. We are adding EG2-based text search alongside our existing search methods.
• Model selection for agents. We are updating our LGND Geo skills so that the AI agents our customers use can choose the most suitable model for each question.

If you'd like to see how this performs on your own use case, send us a concept, a region or a question you're working on, or run it through LGND Geo yourself.