EmbeddingGemma 2 — Text + Image · 440M

A smaller deployment checkpoint derived from Google DeepMind's EmbeddingGemma 2. This export removes the audio encoder and audio projection. It preserves every retained BF16 tensor exactly: there is no fine-tuning, distillation, quantization, or change to the embedding projection.

The 270M/440M names are rounded upstream deployment sizes. File size is not total runtime memory.

Quick start

pip install "transformers>=5.19.0" "sentence-transformers>=6.1.0"

import torch
from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "jayyun98/embeddinggemma-2-text-image-440m",
    device="cpu",
    model_kwargs={"dtype": torch.float32},
)

queries = model.encode(
    ["What causes the northern lights?", "로컬 코드 검색 모델을 찾고 싶어요."],
    prompt_name="SearchQuery",
    normalize_embeddings=True,
)
documents = model.encode(
    ["The northern lights are caused by charged particles from the sun."],
    prompt_name="Document",
    normalize_embeddings=True,
)
print(model.similarity(queries, documents))

FP32 CPU inference was tested with Transformers 5.19.0, SentenceTransformers 6.1.0, and PyTorch 2.14.1. The weight files remain BF16. Upstream supports BF16/FP32 inference and advises against FP16; device-specific BF16 performance was not measured here.

Image and mixed input embeddings

from PIL import Image

image = Image.open("photo.jpg").convert("RGB")
image_vector = model.encode({"image": image}, normalize_embeddings=True)
mixed_vector = model.encode(
    {"text": "A description of this photo.", "image": image},
    normalize_embeddings=True,
)

Image-only, text + image, and interleaved two-image inputs were tested. Text queries and images use the same 768-dimensional embedding space. Use matching dimensions when comparing embeddings from either exported variant or the upstream model.

Code search and smaller vectors

code_queries = model.encode(
    ["Find a Python function that sorts a list."],
    prompt_name="CodeRetrieval",
    truncate_dim=256,
    normalize_embeddings=True,
)
code_documents = model.encode(
    ["def sorted_copy(items): return sorted(items)"],
    prompt_name="Document",
    truncate_dim=256,
    normalize_embeddings=True,
)

For documents with a title, format title: {title} | text: {content} manually and omit prompt_name="Document". Re-normalize truncated vectors and use the same dimension for queries and documents. Original task prompts, mean pooling, and normalization modules are preserved.

Conversion and verification

• Source revision: 914f7f89142e33e77833254d9c9b90c3cef7303b.
• Config changes: audio_config=null.
• Retained weight prefixes: language_model., vision_tower., embed_vision..
• SentenceTransformers modality configuration targets text/image inputs and structured messages.
• Upstream processor/tokenizer assets are retained for standard Transformers compatibility; preprocessing metadata does not restore any removed encoder weights.
• Every exported tensor was checked for exact equality with the source checkpoint.
• Exported models load with exactly the expected runtime state keys and no removed encoders.
• English/Korean search, document and code prompts, plus 128/256/512-dimensional normalized vectors were compared against the full upstream model.
• Image-only, mixed text/image, and interleaved two-image inference were also compared.
• Maximum absolute embedding difference in the tested CPU FP32 cases: 2.98023224e-08.

See conversion.json and verification.json for measurements. These are equivalence and loading checks on a small fixture set, not a new retrieval benchmark. No benchmark suite or hardware speedup was measured. Audio is unavailable; video workflows were not validated for this text/image package.

Attribution

Original model, weights and tokenizer/processor assets: Google DeepMind. This repository is an independent derivative export, not an official Google release. See LICENSE, NOTICE, and the upstream model card.