The embeddinggemma-2 models are a family of embedders that produce vector representations of text for tasks like semantic search, clustering, and personalizing recommendations. They can be loaded and used entirely inside Strata, which runs the model on your own CPU or GPU without sending data anywhere. This makes embedding generation private and fast enough for everyday use on a laptop.
Why embeddinggemma-2 exists
The embeddinggemma-2 models were created to give developers a strong, openly available way to generate high-quality embeddings without depending on hosted services. They come in several sizes so you can pick a model that fits your hardware and accuracy needs.
How to run it on your own machine
Strata bundles the embeddinggemma-2 weights and the inference code, so on your machine you open the Strata app, choose the embeddinggemma-2 variant, and let the app load the model into memory. The Strata engine runs the computation locally; nothing is uploaded to the internet. You can check that the model is loaded by looking at the model status line in Strata while it generates an embedding.
| Variant | Approximate model size | Typical hardware recommendation |
|---|---|---|
| embeddinggemma-2-small | compact enough for consumer laptops | laptop CPU or midrange GPU with at least 8 GB RAM |
| embeddinggemma-2-medium | larger parameter count than small | laptop with a modern GPU and 16 GB+ RAM |
| embeddinggemma-2-large | largest of the series, more compute |
To verify the model is running, you can request an embedding for a short phrase in Strata and watch the generated vector appear in the output panel. Measuring the latency: run the same prompt three times and record the time on each run. The time includes the time the model spends computing; the Strata engine reports the wall-clock duration in seconds.
Embeddinggemma-2 is an embedding model, meaning it takes text and outputs a fixed-size numeric vector that captures meaning.
Checking that it stays local
Strata does not expose any API endpoints to the internet when you use local models, and the Strata engine reads the model files from your disk only. You can confirm privacy by ensuring the app is launched from your machine and not calling external services. If you want extra confidence, read the process list while generating an embedding; no network traffic should appear.
How to measure accuracy yourself
The models are open, so you can test quality by asking the model to embed a pair of sentences that are similar or dissimilar and then comparing the resulting vectors with a simple cosine similarity calculation done in a notebook. Run the same sentence through the model twice, compute the cosine similarity between the two vectors, and compare it to a reference embedding from a different model. This hands-on check shows whether the embeddinggemma-2 size choice gives you the semantics you need.