EmbeddingGemma 300M API

Turn text into vectors for search, clustering and retrieval from $0.001 per million input tokens.

Vector width 768
MTEB score 61.15
Context 2K tokens
$0.001 per million

from openai import OpenAI

client = OpenAI(
    base_url="https://api.modelsheep.com",
    api_key=API_KEY_MODELSHEEP,
)

response = client.embeddings.create(
    model="google/embeddinggemma-300m",
    input="The quick brown fox jumps over the lazy dog.",
)

print(response.data[0].embedding)

How to use the EmbeddingGemma 300M API

1
Create a ModelSheep account and copy an API key from the dashboard.
2
Point the OpenAI SDK at https://api.modelsheep.com and set the model to google/embeddinggemma-300m.
3
Call /embeddings with your text split into chunks that fit the context window, because a longer input loses its tail, and each chunk returns one vector of 768 values.

What EmbeddingGemma 300M is

EmbeddingGemma 300M is Google's small open embedding model. It builds on Gemma 3, reads text in more than 100 languages, and returns a vector 768 numbers wide. Google trained it to run on a phone or a laptop, so it holds 0.308B parameters and reads 2K tokens, the shortest context of any embedding model we serve. It costs less per million input tokens than any other model in our catalogue.

What it does well

Embed short text chunks for a retrieval augmented generation (RAG) index in
Index a large corpus for semantic search when price per million tokens sets
Group support tickets, reviews, or log lines by meaning for clustering and

What it is not for

Long text. The context window is the shortest of the embedding models we
Highest recall. This model scores above BGE M3 on MTEB multilingual v2 but

Calling it

Call POST /embeddings with model set to google/embeddinggemma-300m, and send one string or an array of strings. Google trained the model on a task prefix, so write a query as task: search result | query: {content} and a stored document as title: none | text: {content}. A call without the prefix returns a weaker vector. Set dimensions to truncate the vector to 512, 256 or 128 numbers, which saves index memory. Google's card reports a lower MTEB mean at each of those widths than the 61.15 we publish for 768, and a lower mean again for the quantized checkpoints.

google/embeddinggemma-300m /embeddings

SDKs and languages

EmbeddingGemma 300M answers on an HTTP endpoint, so any language that can post JSON reaches it. These are the ones the box above writes out.

OpenAI SDK: PythonTypeScriptGoJavaC#PHPRust
HTTP: cURLPythonTypeScriptGoRustJavaC#PHP

Specifications

Author	Google
Category	embedding
Accepts	text
Returns	vector
Total parameters	0.308B
Active parameters	dense, all of them
Score	61.15 on MTEB multilingual v2
Languages	100+Google writes trained on, not supports, and publishes no list. It says quality varies with how much training data a language had.
Precision served	FP8
Context	2K tokens
Vector	768 dimensions
License	Gemma

Price

Input

$0.001

per million tokens

Output

n/a

this model returns a vector

Against OpenRouter

50%

below the cheapest of 1 provider

The cheapest OpenRouter price for this model was $0.002 in and n/a out, with a median of $0.002 and n/a across the comparable providers. Read on 2026-09-17.

License

Google releases the weights under the Gemma license. This is not a standard open license. It permits commercial use. Google claims no rights in the vectors you generate. Anyone who passes the weights or a derivative to a third party must hand over the Gemma Terms of Use, bind that party to the use restrictions, and follow the Gemma Prohibited Use Policy. Google can restrict access it believes breaks the terms. Read the terms before you ship.

Parameters: 7.953B
Context: 33K
Vector: 4096
Input per 1M: $0.031
Output per 1M: n/a
License: OpenMDW 1.1