Gemma 4
No published training cutoff
google/gemma-4-31B-it
$0.09 in, $0.34 out per 1M tokens
Small open on-device embedding model for multimodal search, RAG, classification and clustering.
Model idgoogle/embeddinggemma-2
PricePrice not published
Checked7 October 2026
No newer embedding model in our registry yet.
2 ways to use EmbeddingGemma 2, with the model id each provider expects.
| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Hugging Face | google/embeddinggemma-2 |
huggingface.co | |
| Google docs | ai.google.dev |
Price not published.
No price source linked. Checked 7 October 2026.
What its maker marketed at launch, and what people found later, each with its source.
Maps text (including code), images, video and audio into one 768-dimensional space with 740M parameters in total: a 270M text model (130M backbone + 140M embedder) plus optional 170M vision and 300M audio encoders that can be loaded separately.
Embeddings can be cut to 512, 256 or 128 dimensions (up to 6x less storage); quality loss is small down to 256d.
MTEB code v1 78.68 vs 68.76 for EmbeddingGemma 1; MTEB multilingual v2 61.36 vs 61.15; MMEB v2 overall 59.01 (Google’s numbers).
Announced 2026-10-06 on the Google blog. Apache 2.0 (Gemma 4 license), not gated on Hugging Face. Runs on phones and laptops; uses task-instruction prefixes (e.g. ‘task: search result | query: ...’) for text inputs. sentence-transformers id: google/embeddinggemma-2. No hosted Gemini API endpoint is listed; for hosted multimodal embeddings Google offers gemini-embedding-2.
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("google/embeddinggemma-2")
q = model.encode("What causes the northern lights?", prompt_name="SearchQuery")
d = model.encode("The northern lights are caused by charged particles from the sun.", prompt_name="Document")
print(model.similarity(q, d))
Sources: Hugging Face model card, launch blog, docs.