Google releases EmbeddingGemma 2 to find text, images and audio

Google released EmbeddingGemma 2, a small open model that turns text, code, images, video and audio into groups of numbers. Tools that find items use these groups, the embeddings, to find items that match. The model has the Apache 2.0 license, operates on phones and laptops, and is on Hugging Face. It uses Gemma 4.

How it works

Developers install only the encoders (model parts) that they must have. The model has a maximum of 740 million parameters for all media. A method, Matryoshka learning, cuts each embedding to 128 dimensions. This gives 6 times less storage.

The proof

At 128 dimensions, the model keeps about 90% of text quality. For images, video and voice tasks it keeps about 75%. Google tells that the model gives results that are much more accurate than the first version in tasks to find code. Google gave no other numbers.

What is not known

One Reddit user tells that for text only the increase is not large. A different user tells that task prefixes are not optional. The user must cut an embedding first and then set its length to 1. These are words of users, not tests.

Insight: Google gives away the parts that developers put in agents

Google released EmbeddingGemma 2 with the Apache 2.0 license. On 18 October, Google Research released Harness2, agent control software, with the same license. Google also added Gemini Agent to Gemini Enterprise. This reads as a plan to make free Google parts the parts that developers select first, and Google agent tools the easy next step.