EmbeddingGemma 300M

EmbeddingGemma is Google's small text embedding model, announced on 4 September 2025. It does not generate text: it converts a sentence or a document into a vector, which is the foundation of search, retrieval-augmented generation, classification and clustering. The model was built from Gemma 3 with a T5Gemma initialisation and trained on data in more than a hundred languages — but its whole point is size. At 308 million parameters it fits, after quantisation-aware training, into under 200 MB of RAM, and produces an embedding of 256 input tokens in less than 15 milliseconds on an EdgeTPU. That means a search index and a RAG pipeline that live entirely on a phone, with the text never leaving the device. The output vector is 768 numbers long, but thanks to Matryoshka Representation Learning it can be truncated to 512, 256 or 128 and re-normalised — the shorter the vector, the cheaper the index and the faster the search. Google publishes exactly what that costs: on the multilingual MTEB benchmark the score falls from 61.15 at 768 dimensions to 58.23 at 128, on the English set from 69.67 to 66.66, and on the code set most steeply, from 68.76 to 62.96. Quantisation costs strikingly little by comparison — the four-bit checkpoint scores 60.62 in multilingual MTEB against 61.15 for the full-precision model. Google's claim is that this is the highest-ranking open multilingual embedding model under 500 million parameters on the MTEB leaderboard, and comparable with models nearly twice its size. The limits are clear: the context window is 2,048 tokens, so long documents have to be split into chunks, and the activations do not support float16 — bfloat16 or float32 is required. The weights are published under the Gemma terms, which allow commercial use but require accepting the licence on Hugging Face and passing the use policy on downstream. In Google's own recommendation the division of labour is simple: EmbeddingGemma for what runs on the device, the Gemini Embedding API for what runs in the cloud.

▸Videos

No videos yet.

▸Reviews