EmbeddingGemma 2 searches text, images, audio, and video on the device

Important

Lab claim

Google DeepMind released a 740-million-parameter open model that embeds text, images, audio, and video on a device.

How to read this claim

Measurement — The MTEB Code move and the memory figures are Google's. They have not been independently rerun.
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — The MTEB Code move and the memory figures are Google's. They have not been independently rerun.

Keep the habit · field card with prices and worked checks

What changed

Search over private photos, audio, and code usually meant an upload. EmbeddingGemma 2 is a 740 million parameter model that can do that search on the device that already holds the files.

What it introduces

Search over text, images, audio, and video can run on the device that already holds the files, with a 740 million parameter model.

Who benefits Engineers building on-device search, local retrieval for agents, and anyone indexing a codebase or a media library without sending it out. Product teams whose private-data features still assume a cloud upload.

Try this

If a product uploads private files only so they can be searched, test whether this model can do that search on the machine that already holds them.

What we do not know

The MTEB Code move and the memory figures are Google's. They have not been independently rerun.

Evidence

Primary Google DeepMind. Reported by Google DeepMind; not independently verified.

Independent The MTEB Code move from 68.76 to 78.68, and the memory figures, are Google's. They have not been independently rerun.

Sources

Google DeepMind Primary · October 6, 2026
Weights

Still unclear? · Start here · ranks, prices, unknown The MTEB Code move and the memory figures are Google's. They have not been independently rerun.

The record

What happened

On October 6, 2026, Google DeepMind released EmbeddingGemma 2, built on the Gemma 4 architecture, under Apache 2.0. The full model has 740 million parameters. Text-only use can drop to about 270 million, with optional vision and audio encoders. It maps text, code, images, audio, and video into one vector space. Weights are on Hugging Face and Kaggle.

Why it matters

Embeddings are how software finds a file, a clip, or a function without an exact keyword. Doing that on the device means a private library does not have to be uploaded to be searchable. That is a deployment change, not only a leaderboard change. Judge it on whether a laptop agent finds the right local file. The benchmark claims, including “best in class for its size,” are Google’s.

Output vectors can be truncated from 768 dimensions to 512, 256, or 128, so a local index can shrink by up to 6× if you store 128 dimensions instead of 768. Context is 8,000 tokens, four times EmbeddingGemma 1, which Google says covers about 5.5 minutes of audio, 29 images, or 58 video frames. An app that only embeds text does not have to load vision and audio.

Google says that with quantization, a Pixel 11 Pro needs about 191 MB of active RAM for the text-only weights and about 567 MB for the full multimodal model. On code, Google reports MTEB Code moving from 68.76 to 78.68. The lab says the model leads sub-1-billion multimodal embedders on suites such as MTEB Code and the Massive Audio Embedding Benchmark, and that it matches EmbeddingGemma’s multilingual text scores. Those comparisons are the lab’s.

Serving paths named in the post include transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, and LM Studio. Google says the model shares a text tokenizer and audio encoder with Gemma 4, so the embedder and a generative Gemma 4 model can sit in one pipeline. Fine-tuning notes point to Unsloth. Vector storage points to Qdrant. The text-only predecessor passed 20 million downloads by Google’s count. On-device examples in the post are Instant Media Search, Video Moments Finder, and a pairing with Gemma 4 in the AI Edge Foresight app.