yan2026/embeddinggemma-2-GGUF

EmbeddingGemma 2 — GGUF

GGUF conversions of google/embeddinggemma-2 for llama.cpp.

Multimodal embedding model (text + vision + audio) by Google. Output dimension: 768.

Files

File	Type	Size
google-embeddinggemma-2-bf16.gguf	Main model (bf16, lossless from original weights)	533 MB
google-embeddinggemma-2-q8_0.gguf	Main model (q8_0)	290 MB
google-embeddinggemma-2-q4_K_M.gguf	Main model (q4_K_M)	174 MB
mmproj-google-embeddinggemma-2-bf16.gguf	Multimodal projector (vision + audio, bf16)	937 MB

Usage

Requires a recent llama.cpp build (2025-10+ / commit 8bb3cf0d2 or newer) with EmbeddingGemma 2 support.

llama-server (OpenAI-compatible embeddings)

llama-server \
    -m google-embeddinggemma-2-q8_0.gguf \
    --mmproj mmproj-google-embeddinggemma-2-bf16.gguf \
    --embeddings \
    -ngl 99 -b 2048 -ub 2048 \
    --port 8080

curl http://localhost:8080/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{"model": "embeddinggemma-2", "input": ["your text here"]}'

Task instruction prefixes (important for quality)

Per the original model card, queries and documents should be prefixed by task:

Use case	Query	Document
Web / document search	task: search result | query: {query}	title: {title} | text: {content}
Question answering	task: question answering | query: {question}	title: {title} | text: {passage}
Fact checking	task: fact checking | query: {claim}	title: {title} | text: {evidence}
Symmetric similarity	raw text	raw text

Use title: none when no title is available.

Notes

• Embeddings are L2-normalized by llama.cpp by default: dot product = cosine similarity.
• The llama.cpp /v1/embeddings endpoint currently supports text-only embeddings; vision/audio encoders require the mmproj file and are for multimodal use (audio support is experimental upstream).
• Converted with convert_hf_to_gguf.py from the official google/embeddinggemma-2 release (commit 914f7f8).

License

Gemma Terms of Use — see license link. Original model © Google.