theTNai
Get the brief

Open source · Google · 6 Oct 2026

Google released EmbeddingGemma 2, an open 740-million-parameter model that turns text, code, images, audio and video into searchable vectors.

It is under the Apache 2.0 licence and is on Hugging Face and Kaggle. Google says text-only use needs as few as 270 million parameters, and it reports the full model running on a Pixel phone in about 567 MB of memory.

What Google released

EmbeddingGemma 2 is an embedding model. It turns text, code, images, audio and video into lists of numbers in one shared space, so a program can find items that mean similar things. A voice memo can find a video clip, and a typed query can search hours of audio. It is built on the Gemma 4 architecture.

The full model has 740 million parameters, and its parts are modular. Text-only use needs as few as 270 million. A vision encoder adds 170 million and an audio encoder adds 300 million, and both are optional.

Size, memory and limits

The model reads up to 8K tokens at once, four times the first EmbeddingGemma. In one input it can take up to 5.5 minutes of audio, 29 images or 58 video frames, alone or mixed. Its vectors have 768 dimensions and can be cut to 512, 256 or 128. Google says this saves up to 6 times the storage and memory in a local vector database.

With quantisation, Google reports about 191 MB of active memory for text-only use and about 567 MB for the full model on a Pixel 11 Pro phone.

How it compares and what it is for

Google says its multilingual text results match the first EmbeddingGemma, and its MTEB Code score rose from 68.76 to 78.68. It also claims top scores among multimodal embedding models under 1 billion parameters on MTEB Code and MAEB. These are the maker's own figures. FoneArena reports that the first EmbeddingGemma passed 20 million downloads.

Google suggests it for offline search across media, private retrieval-augmented generation (RAG) with Gemma 4, and code search for coding agents. It shares Gemma 4's text tokenizer and audio encoder, so the two can run together with less memory.

How to try it

  1. Download the weights from Hugging Face or Kaggle. On-device versions are in the LiteRT Community on Hugging Face.
  2. Run it in Python with transformers or sentence-transformers, or locally with Ollama, LM Studio, llama.cpp or MLX. vLLM and SGLang are listed for serving.
  3. For phone apps, use Google AI Edge MediaPipe or LiteRT. In a web browser, use transformers.js or WebGPU.
  4. Store the vectors in a vector database such as Qdrant, which Google lists as supported.

How to get good results

Our practical advice, based on what the sources below say about it.

  1. Load only the parts you needFor text or code search, skip the vision and audio encoders. Text-only use needs as few as 270 million parameters and about 191 MB of memory when quantised.
  2. Shorten vectors to save spaceVectors can be cut from 768 to 512, 256 or 128 dimensions. Google says this saves up to 6 times the storage. Compare search quality at each size on your own data before you choose.
  3. Stay within the input limitsEach input can be up to 8K tokens, 5.5 minutes of audio, 29 images or 58 video frames. Split longer files into parts before embedding them.
  4. Pair it with Gemma 4 for private RAGIt shares Gemma 4's text tokenizer and audio encoder, so running both together uses less memory. That suits a question-answer system that keeps your files on your own device.
  5. Fine-tune if results are weakIf general results are poor on your own documents, Unsloth has published guidance for fine-tuning the model.

Why it matters in Tamil Nadu

Students and startups can build search over their own documents or photos on a laptop or phone, without paying for an API.

Where this comes from

  • SourceGoogle
  • Also checkedYourStory Licence, context window, media limits and where to download
  • Also checkedFoneArena Parameter split, memory use, vector sizes, benchmarks and supported tools
  • Checked against the source11 Oct 2026
  • Written bytheTN.ai's AI agent, in its own words
  • Approved byPaul Jeevanesan, Founder and Chief EditorAlso founder and CEO of StudAI One, whose services wing is StudAI Works (related parties, see disclosures).

Spotted an error? Tell us: we fix confirmed errors within 24 hours and log them in public.

Cite this page

theTN.ai, "Google released EmbeddingGemma 2, an open 740-million-parameter model that turns text, code, images, audio and video into searchable vectors.", 6 Oct 2026. https://thetn.ai/tech/2026-10-06-google-embeddinggemma-2/

More in Open source