What it is
EmbeddingGemma 2 is an open embedding model from Google DeepMind, released on October 6, 2026 under Apache 2.0. Embedding models turn content into vectors so software can find related material for search and retrieval-augmented generation (RAG). Built on Gemma 4, it maps text, code, images, video and audio into one 768-dimensional space, in more than 100 languages.
It has 740 million parameters in modular parts: a 270-million-parameter text model, a 170-million vision encoder and a 300-million audio encoder, and encoders you switch off are never loaded. Google says the quantised model needs about 191 MB of active RAM for text and about 567 MB for everything on a Pixel 11 Pro. Inputs share an 8,192-token window, enough for about 29 images, 58 video frames or 5.5 minutes of audio, and vectors can be cut to 512, 256 or 128 dimensions.
Our review
EmbeddingGemma 2 is small enough to index a phone's photos, recordings and documents in one vector space without a network call, and that is its whole argument. The closed alternative, Google's own Gemini Embedding 2, does the same across media but bills per token, from $0.20 per million for text to $12 for video. This model does it in under 600 MB of RAM under Apache 2.0.
The quality gains are concentrated in code. Google reports MTEB Code rising 9.92 points to 78.68, while multilingual text barely moves, from 61.15 to 61.36. That text score is respectable for the size but short of larger open models: Qwen reports 64.33 for its 0.6-billion-parameter Qwen3-Embedding and 70.58 for the 8B on the same benchmark. The card's image, video and audio results come without rival comparisons, so Google's claim to lead among sub-1B multimodal embedders is its own.
The engineering choices help small deployments. Text-only apps load 270 million parameters; truncating to 256 dimensions keeps most text quality and about 95% of multimodal quality by Google's figures. At 128 dimensions multimodal search keeps only about 75%, which makes the 6x storage saving a poor trade for photo search.
Two warnings from the card matter more than any benchmark. Running it in float16 overflows and returns NaN or quietly degraded vectors without raising an error, which is easy to miss on GPUs that default to fp16. And the card's licence link points to the Gemma 4 licence page while the text says Apache 2.0, so legal teams will want that settled before shipping.
Where it earns its place
An Android developer building a notes app with voice memos and photos
Let users search everything they have saved by typing a phrase, with nothing leaving the phone
Ship the LiteRT build and index on-device; Google's Pixel 11 Pro figures put the full model at about 567 MB of RAM. There is no per-query cost, but expect about 75% of multimodal quality if you truncate to 128 dimensions to save storage.
An engineer at a 50-developer company running code search on a CPU-only internal server
Replace keyword search over internal repositories with semantic search
Load text-only (270M parameters) in fp32 or bf16 with sentence-transformers and use the CodeRetrieval prompt. MTEB Code of 78.68 is the model's strongest score; a million 768-dimension vectors take about 1.5 GB in bf16.
A privacy lead at a telehealth provider
Add semantic search over consultation recordings without sending audio to a cloud API
Self-hosting removes per-token fees and third-party processing. Audio must be mono 16 kHz and each input holds about 327 seconds, so recordings need chunking, and Google publishes no audio-benchmark comparison with rivals.
Before you commit
- Setup
- Load from Hugging Face with sentence-transformers 6.1 or later, or pull via Ollama or LMStudio. Text queries need task prefixes such as 'task: search result | query:' for best precision.
- Integrations
- Transformers, sentence-transformers, LiteRT and MediaPipe for on-device, transformers.js on WebGPU, Ollama, vLLM, SGLang, MLX, llama.cpp and Qdrant; Unsloth for fine-tuning.
- Data and privacy
- Weights run entirely on your hardware, so no data leaves it. The card states Apache 2.0 but links to the Gemma 4 licence page, and use must follow the Gemma Prohibited Use Policy.
- Cost at scale
- No licence fees; the cost is storage and compute. A million 768-dimension vectors take about 1.5 GB in bf16, or about 250 MB at 128 dimensions, and the text-only model is small enough for CPU inference.
- Limits
- 8,192-token shared input window (about 29 images, 58 video frames at 1 fps or 327 seconds of audio). Do not run in float16. Re-normalise after truncating, and use the same dimension for queries and documents.
What's good
- Text, code, image, video and audio in one 768-dimension space
- About 191 MB active RAM for text-only on a Pixel 11 Pro
- Unused encoders never load; vectors truncate to 512, 256 or 128
What's not
- Multilingual text quality unchanged from v1 (61.36 against 61.15)
- Silently corrupts output in float16
- Multimodal benchmarks lack rival comparisons; licence link ambiguous
Best for
Mobile and edge developers building private or offline search across text, code, images and audio.
Alternatives
EmbeddingGemma 2: common questions
- Is EmbeddingGemma 2 worth it?
- EmbeddingGemma 2 makes private, offline search across photos, voice notes and code practical on a phone, while leaving text-only retrieval quality almost exactly where version one left it. We rate it 4.3 out of 5.
- How much does EmbeddingGemma 2 cost?
- Free under Apache 2.0, subject to the Gemma Prohibited Use Policy. No hosted API from Google or Hugging Face inference providers yet.
- What are the downsides of EmbeddingGemma 2?
- Multilingual text quality unchanged from v1 (61.36 against 61.15); silently corrupts output in float16; multimodal benchmarks lack rival comparisons; licence link ambiguous.
- Who is EmbeddingGemma 2 best for?
- Mobile and edge developers building private or offline search across text, code, images and audio.
- How much does EmbeddingGemma 2 cost for a team or in production?
- No licence fees; the cost is storage and compute. A million 768-dimension vectors take about 1.5 GB in bf16, or about 250 MB at 128 dimensions, and the text-only model is small enough for CPU inference.
- Is EmbeddingGemma 2 safe for company data?
- Weights run entirely on your hardware, so no data leaves it. The card states Apache 2.0 but links to the Gemma 4 licence page, and use must follow the Gemma Prohibited Use Policy.
- What are the best EmbeddingGemma 2 alternatives?
- The closest alternatives we have reviewed are Gemini Embedding 2, Qwen3-Embedding and Qwen3-Reranker and all-MiniLM-L6-v2.
Answers checked Oct 22, 2026.
Make EmbeddingGemma 2? Show this rating on your site
