EmbeddingGemma 2 at a glance
Google released EmbeddingGemma 2 on October 6, 2026: an open multimodal embedding model on the Gemma 4 architecture that maps text, code, images, video and audio into one 768-dimension embedding space. Its 740M parameters load modularly, the licence is Apache 2.0, and the weights are free on Hugging Face and Kaggle.
Ask your AI about this story
Ask what changed in EmbeddingGemma 2 and what to check before using it in your business. Compare it on price, speed, intelligence and the price-to-speed ratio against the models I would otherwise use, and say which tasks it suits or does not suit.
What changed with EmbeddingGemma 2
On October 6, 2026, Google released EmbeddingGemma 2, an open multimodal embedding model built on the Gemma 4 architecture. It maps text, code, images, video and audio into a single 768-dimension embedding space, so a voice memo can find a video clip or a text query can search a photo library. It has 740 million parameters in a modular design: a roughly 270M-parameter text-only base, plus optional 170M vision and 300M audio encoders you load only if you need them. Its 8,192-token context window fits about 5.5 minutes of audio, 29 images or 58 video frames per input. The licence is Apache 2.0, not the Gemma Terms of Use, and the open weights are free on Hugging Face and Kaggle, with day-one support in sentence-transformers, llama.cpp, vLLM, SGLang, Ollama, LM Studio and LiteRT/MediaPipe. Gemini Enterprise Agent Platform Model Garden availability was announced as coming soon.
The benchmark scores are Google's own, from the model card: 78.68 NDCG@10 on MTEB Code, up from 68.76 for the first EmbeddingGemma, a mean of 61.36 on MTEB Multilingual v2, and 59.01 overall on MMEB v2 at the full 768 dimensions. Matryoshka Representation Learning lets you truncate the vectors from 768 down to 512, 256 or 128 dimensions, cutting vector storage by up to 6x. The model is built to run on device: Google puts its quantized footprint at about 191MB of RAM in text-only mode and about 567MB with the full multimodal stack on a Pixel 11 Pro. As of November 17, 2026, Artificial Analysis had not published a page for EmbeddingGemma 2 and Arena does not list it, so there are no independent scores yet. Google published the model's technical report on arXiv on November 17, 2026.
EmbeddingGemma 2 in numbers
Values are copied from the linked source on the date shown. They are not IA Compa measurements.
- MTEB Code NDCG@10 (Google's model card)as of · Source for MTEB Code NDCG@10 (Google's model card): ai.google.dev (opens in a new tab)
- 78.68
- MMEB v2 overall at 768 dimensions (Google's model card)as of · Source for MMEB v2 overall at 768 dimensions (Google's model card): ai.google.dev (opens in a new tab)
- 59.01
- Context windowas of · Source for Context window: ai.google.dev (opens in a new tab)
- 8,192 tokens
How it compares
Current models we have covered: released in the last six months, newest version of each line only. Each value is the model's latest recorded value. (Models)
Where to compare EmbeddingGemma 2
Independent leaderboards and official resources for comparing this model yourself. Rankings change often; each site shows its own current data.
General leaderboards
Why EmbeddingGemma 2 matters for businesses
Why it matters: embeddings are the plumbing behind search and retrieval, and this release lets a business run that plumbing itself. A firm can index its documents, code, product photos and call recordings in one space and search across all of them on its own hardware, with no per-query API bill and no customer data leaving the building. Apache 2.0 means you can ship it inside a commercial product without negotiating a licence, and the modular encoders plus Matryoshka truncation keep the phone, laptop or server footprint, and the vector database bill, small.
Put it to work: AI services
Get AI News by email
Verified AI stories, each with its official source, on Monday and Thursday mornings. Add breaking news to hear about the biggest ones as soon as we publish them.
Every story links its official source
Free, in English or Spanish
Unsubscribe with one click, any time
Sources for this story
- Introducing EmbeddingGemma 2: On-device, open multimodal embeddings (opens in a new tab)Official
Google · blog.google
- EmbeddingGemma 2 model card (opens in a new tab)Official
Google · ai.google.dev
- google/embeddinggemma-2 on Hugging Face (opens in a new tab)Official
Google on Hugging Face · huggingface.co
- EmbeddingGemma 2 technical report (arXiv:2610.37364) (opens in a new tab)Official
Google (arXiv) · arxiv.org
Last verified: . IA Compa summarizes public announcements for business readers. We are not affiliated with the companies mentioned. Always confirm details, pricing, and availability with the official source.
- #embedding-model
- #multimodal
- #open-weights
- #on-device
- #gemma
- #semantic-search

