EmbeddingGemma 2 is out! Our first native multimodal embeddings built on Gemma 4 under Apache 2.0. 🤗 Embeds 100+ languages, code, images, audio, and video into one vector.
🧮 8,192 context
3️⃣ 740M omni, 440M text+vision, 570M text+audio, or 270M text-only
📈 +14% jump on MTEB Code
🪆 Matryoshka: 768d, 512d, 256d, or 128d
🤗 Available in Sentence Transformers and LiteRT-LM
🔓 Apache 2.0
huggingface.co/google/embeddi…
🧮 8,192 context
3️⃣ 740M omni, 440M text+vision, 570M text+audio, or 270M text-only
📈 +14% jump on MTEB Code
🪆 Matryoshka: 768d, 512d, 256d, or 128d
🤗 Available in Sentence Transformers and LiteRT-LM
🔓 Apache 2.0
huggingface.co/google/embeddi…
Oct 6, 2026 · 6:33 PM UTC
22
61
575
33,171
Try directly in your browser: huggingface.co/spaces/webml-c…
1
0
8
1,838
@_philschmid Any support for windows yet ?
We need the Omni experience 😬
We need the Omni experience 😬
0
0
1
133
@_philschmid tyvm cant wait to make more babygemmas thank you for listening to our feedback. your delivery rate on suggestions is faster than microwaved pizza, seriously
0
0
0
11
@_philschmid The 128d option makes a nice two-stage setup: shortlist with 128d vectors, then rescore the shortlist at 768d. For web RAG the bigger win is upstream. Strip nav and footers before embedding, or every page from one site starts to look alike. Notes on that: agentsearchhq.com/blog/url-to-ma…
0
0
1
90
@_philschmid One vector for truly multimodal retrieval is a huge step forward
0
0
1
106
@_philschmid The Matryoshka support for storage savings is such a clever way to handle these embeddings. We actually went deeper on this here: x.com/AGTPinsights/s…
Google just released EmbeddingGemma 2, an open multimodal embedding model that runs on-device. Here's what you need to know.
Sundar Pichai announced it as Google's first open, natively multimodal embedding model. It maps text, code, images, video and audio into one shared lows the original EmbeddingGemma, which passed 20 million downloads.
It is built on the Gemma 4 architecture and released under an Apache 2.0 license, so commercial use is allowed. Weights are live on Hugging Face now. It supports 100+ languages and Sentence Transformers.
The design is modular. You can load only the encoders you need. Text and code run on a 270M model. Adding vision (170M) and audio (300M) brings the full model to 740M parameters. Embeddings are 768-dimensional, and Matryoshka support lets you truncate them to 128 dimensions for about 6x storage savings.
It is built for offline, private use. You could find a moment in a video using a voice memo, or search hours of audio with a text query, with no internet connection. Paired with Gemma 4, it powers on-device RAG: EmbeddingGemma 2 retrieves local files and Gemma 4 reasons over them, with lower memory use.
Performance: it is best in its size class on MTEB (Code) and MIEB (Lite), and beats some specialist models more than twice its size. Google cites a 9.92-point code improvement over the previous model.
Day-0 support came from llama.cpp, Unsloth (GGUF) and Transformers.js. Developers report it running fully in the browser on WebGPU at about 20 to 70 ms per query.
Key numbers:
- 740M parameters total (270M text and code, 170M vision, 300M audio)
- Runs in about 0.5GB of RAM, per Unsloth
- 768-dimensional embeddings, down to 128 with Matryoshka
- 100+ languages
- Apache 2.0 license
- 20M+ downloads for the first EmbeddingGemma
Google announced it on October 6, 2026, with open weights available on Hugging Face.
0
0
1
205
@_philschmid Local. 740M params and Apache 2.0 means a small company can finally search its own photos and PDFs without shipping them to anyone's cloud
0
0
0
86
@_philschmid Benchmarked the video side on MSR-VTT vs SigLIP 2, jina-clip-v2 and CLIP. The 100+ languages claim holds up: top-1 in Japanese and Hindi is 35-37% vs 20-23% for SigLIP 2.
x.com/UnFazed213/sta…
x.com/UnFazed213/sta…
I raced Google's new EmbeddingGemma 2 against the best open vision models on 1,000 videos.
In English it's a coin flip.
Switch languages and it's not even close.
0
0
0
5
@_philschmid One vector space for text, code, images, audio AND video — this changes the multimodal RAG game. Matryoshka dims mean you can trade recall vs latency per use case too. Huge release.
0
0
0
@_philschmid EmbeddingGemma 2 offre-t-il une spacialité aux LLM ? Dois-je l'installer dans le contexte où le MCP ? Peut-on le charger avec AIEdge ?
0
0
0
47
