EmbeddingGemma 2
Google · google/embeddinggemma-2
Weights license: Apache 2.0 · code and components may use other terms
Quality
MTEB (Multilingual, v2)
Mean(task) · 0–100 · higher is better · this model outlined
- Color = publisher
Source: MTEB results dataset (independent runs, CC0) · publisher-reported where the model card is the only source · other benchmarks and versions are not mixed in
| MTEB (Multilingual, v2) | mean(task) · 768d | 61.36 | publisher-reported |
|---|---|---|---|
| MTEB (Multilingual, v2) | mean(task) · 256d | 60.41 | publisher-reported |
| MTEB (Multilingual, v2) | mean(task) · 128d | 57.89 | publisher-reported |
| MTEB (Code) | — · 768d | 78.68 | publisher-reported |
| MIEB-lite | — · 768d | 64.64 | publisher-reported |
| MMEB v2 (image) | — · 768d | 57.28 | publisher-reported |
| MMEB v2 (visdoc) | — · 768d | 67.84 | publisher-reported |
| MMEB v2 (video) | — · 768d | 50.67 | publisher-reported |
| MSEB (retrieval) | — · 768d | 69.54 | publisher-reported |
| MAEB | — · 768d | 49.39 | publisher-reported |
Reported by the publisher
- Massive Text Embedding Benchmark · Mean, Multiple (MTEB, multilingual, v2; Task)
- 61.36Text
- Massive Text Embedding Benchmark · Mean, NDCG@10 (MTEB, code, v1; Task)
- 78.68Text
- Massive Image Embedding Benchmark · Mean, Multiple (MIEB, lite; TaskType)
- 64.64Image
- Massive Multimodal Embedding Benchmark · Mean, Hit@1 (MMEB v2 - Image; Task)
- 57.28Image
- Massive Multimodal Embedding Benchmark · Mean, NDCG@5 (MMEB v2 - VisDoc; Task)
- 67.84Image
- Massive Multimodal Embedding Benchmark · Mean, Hit@1 (MMEB v2 - Video; Task)
- 50.67Video
- Massive Sound Embedding Benchmark · Mean, MRR@10 (MSEB, Retrieval; Task)
- 69.54Audio
- Massive Audio Embedding Benchmark Hugging Face · Mean, Multiple (MAEB; Task)
- 49.39Audio
Values as printed on the card, with the publisher's own settings (card section: Benchmark Results › Overall Evaluation Results (768d)); column “EmbeddingGemma 2”. Not comparable across publishers and not used in any ranking. Model card
Quality × memory
No comparable quality × memory pair
Embedding memory estimates cluster at a few GB, so a quality × memory chart would not separate these models; MTEB is ranked above instead.
Consider EmbeddingGemma 2 for retrieval and semantic search: it turns text, images, video, and audio into 768-dimension vectors.
The publisher's memory figures are for specific setups and are under How to run, with the prompt format; the hardware list uses the planning estimate. PT-BR is not confirmed by a source.
Runs on
Fast memory holds it 119
The planning estimate fits the memory the model can use quickly: graphics card VRAM, or unified memory within the macOS GPU share.
Show all 119
Hardware profiles grouped by the embedding planning estimate (8-bit weights, 20% allowance, 2 GB reserve). Not a measured fit.
Measured speed
No published measurement for this checkpoint
Throughput (texts per second) is shown only when a source publishes it with hardware, build and setup. It is never estimated.
How to run
Published memory: 191 MB (Active RAM, text only, quantized, on-device (Pixel 11 Pro), as stated by Google.); 567 MB (Active RAM, full multimodal, quantized, on-device (Pixel 11 Pro), as stated by Google.). The planning estimate uses 8-bit weights, 20% working allowance and a 2 GB reserve; it is a ceiling, not a measurement.
- Runs in
- Transformers, Sentence Transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LM Studio, Transformers.js, Qdrantsource
- Query/document prompts
- Task-instruction prefixes on text only. Queries: 'task: {task} | query: {content}' (for example 'task: search result | query: {query}'). Documents: 'title: {title} | text: {content}', with 'title: none' when there is no title. Symmetric tasks (classification, clustering, similarity) use the 'task: ... | query: ...' form for all inputs. Images, video and audio take no prefix.
- Languages · PT-BR
- 100+ languagesPT-BR unconfirmed
Documented ways to run it
Copied as printed in the linked source (read 2026-10-18); not run by this site.
pip install -U sentence-transformers transformersfrom sentence_transformers import SentenceTransformer
model = SentenceTransformer("google/embeddinggemma-2")
query = "What causes the northern lights?"
document = "The northern lights are caused by charged particles from the sun."
query_emb = model.encode(query, prompt_name="SearchQuery")
doc_emb = model.encode(document, prompt_name="Document")
print(model.similarity(query_emb, doc_emb))Technical details
What the source and the catalog store for this checkpoint. Missing fields are left out, never zero.
- Checkpoint / source
- google/embeddinggemma-2
- Task
- embeddings for search (feature extraction)
- Library in source
- Transformers, Sentence Transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LM Studio, Transformers.js, Qdrant
- Weight license
- Apache 2.0
- Commercial use
- Allowed Apache License 2.0 text (SPDX) · read 2026-10-18license text
- Parameter basis
- Model card: 740M total = 270M text/code backbone + 170M vision encoder + 300M audio encoder (encoders load separately). The Hub safetensors count is 744M BF16.
- Announced
- 2026-10-06
Ollama's library page for embeddinggemma-2 lists tags 270m, 440m, 570m and 740m and shows a 256K context window for every tag. The model card and Google's blog say 8,192 tokens, and the 440m/570m sizes do not map to the card's 270M/740M description. None of Ollama's figures are used here; verify per-variant specs before publishing them. The Hub tags this repository feature-extraction; the multimodal sub-task is curated from the model card (text, image, video and audio in one vector space). Scores are the publisher's own, from the model card. PT-BR stays unconfirmed (2026-10-07): the model card, the Hub card metadata (language: multilingual) and Google's blog say "100+ languages" without naming Portuguese, and the public MTEB results dataset has no run for this model yet, so there is no Portuguese-subset score. EmbeddingGemma 1's Portuguese result does not carry over to this model.
Sources
Identity and published specifications come from the checkpoint page. Benchmark scores belong to the named release and do not measure speed on your computer.
- Publisher model card
- Hugging Face model card · 2026-10-07
- Google blog, 2026-10-06 · 2026-10-07
- Ollama library (variant and context figures not used) · 2026-10-07
Curated facts checked: 2026-10-07. Each value cites its source (the publisher's card or the MTEB results dataset); estimates are labelled as estimates, and a value no source states stays empty. A repository commit is not a model release date.






























































