Dev48
Language
  • About
  • Services
  • Industries
  • Technologies
  • Articles
  • Contacts
Book a call
    Home/Articles/How small can googles new embeddinggemma 2 get
Dev48
info@dev48.org+375 29 320-40-96WhatsApp@isky_by

© 2026 · All rights reserved.

How Small Can Google's New EmbeddingGemma 2 Get?

Source: Qdrant•October 6, 2026 at 04:01 PM

How Small Can Google's New EmbeddingGemma 2 Get?

Google’s new EmbeddingGemma 2 can run on a phone. The model travels light. The vectors need their own suitcase. At 10 million documents, full-size float32 vectors alone take 30.7 GB of RAM. We got early access ahead of today’s Google DeepMind release and tested how much of that…

Google’s new EmbeddingGemma 2 can run on a phone. The model travels light. The vectors need their own suitcase. At 10 million documents, full-size float32 vectors alone take 30.7 GB of RAM.

We got early access ahead of today’s Google DeepMind release and tested how much of that memory we could save in Qdrant. Quantized full-size vectors kept 99% of their retrieval quality with 30x less vector RAM. Shorter vectors with rescoring reached 77x less vector RAM and retained 94.5%.

We tried both ways of saving memory: shorter vectors and fewer bits per dimension. The chart compares them with exact search over full-size float32 vectors on five text retrieval datasets. Quality is the percentage of the baseline’s nDCG@10 retained, a score for how well the top 10 results rank relevant documents.

Retained nDCG@10 by vector RAM1-bit TurboQuant, % of exact 768d float32 search020406080100float32, 768dbaseline100.030x less RAM, 768dno rescoring99.030x less RAM, 768drescoring99.743x less RAM, 512dno rescoring97.343x less RAM, 512drescoring98.877x less RAM, 256dno rescoring88.177x less RAM, 256drescoring94.5% of nDCG@10 retained (rescoring: oversampling 4)

EmbeddingGemma 2 is an open embedding model built on Gemma 4. It maps text, code, images, video, and audio into one 768-dimensional space. Its text-only path has 270M parameters. Google reports a 14% gain on code retrieval over the first EmbeddingGemma and an 8K-token context window.

Matryoshka Representation Learning lets you keep the first 512, 256, or 128 dimensions without re-embedding your data. Those first dimensions remain useful on their own. Quantization saves more memory by storing each dimension in fewer bits.

Start With Full-Size, 1-Bit Vectors

This is where we’d start: keep all 768 dimensions, use 1-bit TurboQuant, and leave rescoring off. Each vector takes 104 bytes instead of 3,072. That puts the vectors for 10 million documents at about 1 GB of RAM.

Leaving rescoring off also avoids reading original vectors during search. On Quora’s 523k documents, this setup took less than half the query time of float32 search and retained 99.0% of its retrieval quality.

High relevance doesn’t mean identical results. Without rescoring, about 74% of the top 10 results matched exact search. Try rescoring if your application depends on recovering the same neighbors.

Cutting dimensions feels like the obvious way to save memory. In these tests, we’d cut bits first. At 72 bytes per vector, 512 dimensions at 1 bit kept 97.3% of the baseline. At a similar memory budget, the 128-dimensional, 4-bit setup kept 84.0%. Try keeping more dimensions and compressing each one more heavily first.

Fit Into 77x Less Vector RAM With Rescoring

If you’re counting every byte, 256 dimensions is worth a look. Keep 1-bit TurboQuant. Each vector takes 40 bytes in RAM. With rescoring, this setup retained 94.5% of the baseline. Without it, retention dropped to 88.1%. If the full-size setup still needs too much vector RAM, this is worth testing. Check whether the relevance loss and rescoring time fit your application.

Qdrant’s vector datatypes let you store originals as float16 or turbo4 to save disk space. We haven’t tried those options yet. That’s another experiment.

Try It on Your Own Data

Start with the boring configuration: full-size vectors, 1-bit TurboQuant, rescoring off. It did well enough to earn the first test. Compare relevance and latency with an unquantized collection on your own queries. Try rescoring if you need better neighbor recovery. Shorten the vectors if you still need to save memory.

Thanks

Thanks to Google DeepMind for EmbeddingGemma 2 and for early access to the model.

← All articles

What this article says

More in Software Development

All →
Medusa Community Skills for Agents
Medusajs

Medusa Community Skills for Agents

Continental Considers Discontinuing Tire Production at Its Gqeberha Plant in South Africa
Continental

Continental Considers Discontinuing Tire Production at Its Gqeberha Plant in South Africa

Tech Mahindra Marks 40 Years with Bell-Ringing Ceremony at the Bombay Stock Exchange
Tech Mahindra

Tech Mahindra Marks 40 Years with Bell-Ringing Ceremony at the Bombay Stock Exchange

Legacy modernization case studies: 14 projects by trigger
Netguru

Legacy modernization case studies: 14 projects by trigger

The Orchestrators: Inside the Build with Peter Bojtos
Camunda

The Orchestrators: Inside the Build with Peter Bojtos

4 (out of 7) trends and predictions shaping B2B digital commerce in 2027
Commercetools

4 (out of 7) trends and predictions shaping B2B digital commerce in 2027

More from Qdrant

Does a Code-Tuned Embedding Model Pay Off? papr-embed-code-1 in Qdrant
Qdrant

Does a Code-Tuned Embedding Model Pay Off? papr-embed-code-1 in Qdrant

How to Run an Embedding Model on Bare Metal
Qdrant

How to Run an Embedding Model on Bare Metal

Faster and Cheaper Search With Qdrant 1.20
Qdrant

Faster and Cheaper Search With Qdrant 1.20

Scalable PDF Retrieval with Cohere Embed 5 and Qdrant Asymmetric Quantization
Qdrant

Scalable PDF Retrieval with Cohere Embed 5 and Qdrant Asymmetric Quantization

Qdrant 1.20 - Full-Text Search Improvements
Qdrant

Qdrant 1.20 - Full-Text Search Improvements

UltraBERT Fits 32K Tokens in One Vector. RIP Chunking?
Qdrant

UltraBERT Fits 32K Tokens in One Vector. RIP Chunking?