google/EmbeddingGemma 2 740M

• License: Apache-2.0
• Release date: 2026-10-06
• Modalities: Input: text, images, audio; Output: embeddings
• Original model: google/embeddinggemma-2

Description

EmbeddingGemma 2 is a lightweight multilingual embedding model for retrieval, semantic search, routing, and RAG. The upstream model maps text, code, images, video, and audio into a shared embedding space. This registry entry serves its text, image and audio embedding paths through the local llama-server embedding API.

The model produces 768-dimensional vectors and supports Matryoshka reduction to 512, 256, or 128 dimensions. It supports an 8K-token context window.

Run

self serve embeddinggemma-2:740m

Here is a link to the API docs: Embeddings API.

Example Request

With the server running, open another terminal and send a request:

curl http://localhost:8080/v1/embeddings \
	-H 'Content-Type: application/json' \
	-d '{"input":"task: code retrieval | query: How do I index a codebase?"}'

The server returns a JSON response containing the embedding vector.

Engine Parameters

Served by the upstream llama-server through the embedding adapter. These are startup settings, not per-request API fields. Settings not specified in the registry are left to the bundled engine's defaults.

Parameter	Default	What it controls
context_size	8192	Engine context size in tokens; accepts integers from 512 to 1048576.
pooling	mean	Embedding pooling; accepts none, mean, cls, last, or rank.
threads	Engine default (not set in registry)	Number of CPU workers; accepts integers from 1 to 1024.
gpu_layers	Engine default (not set in registry)	Layers offloaded to GPU; accepts integers from -1 to 10000, with -1 meaning all layers.
flash_attn	Engine default (not set in registry)	Flash Attention mode: auto, on, or off; availability depends on the backend and model.
parallel	Engine default (not set in registry)	Number of parallel engine slots; accepts integers from 1 to 256.

Override startup settings with --set <parameter>=<value>.

Quantizations and Files

Default variant: q8. Sizes and SHA-256 hashes are pinned in the model registry.

q8 (default)

File	Role	Size	SHA-256	Origin
embeddinggemma-2-Q8_0.gguf	model	295.6 MiB (309855456 bytes)	2188ac1deca4b77dffefd603c2776a9d76d9d74ec01841392982ebb840b09135	ggml-org/embeddinggemma-2-GGUF
mmproj-embeddinggemma-2-Q8_0.gguf	mmproj	529.1 MiB (554821024 bytes)	c4a8a52691ecef40618438928bdf9e68379b854e24166f292592353db0aab64f	ggml-org/embeddinggemma-2-GGUF

Total download: 824.6 MiB (864676480 bytes).

Notes

• Requires the bundled llama.cpp b11476 runtime or a compatible newer release with gemma-embedding2 support.
• For search queries, use task: search result | query: <query>; for code search, use task: code retrieval | query: <query>. Format indexed documents as title: <title or none> | text: <content>. The adapter does not add these prefixes.
• The API returns the full vector and does not support the dimensions request field. For Matryoshka reduction, truncate and normalize vectors consistently for queries and documents according to the upstream model card.
• The bundled Q8 projector enables image and audio inputs. Use content-object batches documented in the Embeddings API. Media must be inline PNG/JPEG or PCM16 WAV; remote URLs and video are not supported.

Quality benchmarks

MTEB: Publisher-reported

google/embeddinggemma-2

Publisher-reported full-precision checkpoint, 768 dimensions. Not confirmed in the independent MTEB model index; no leaderboard rank.

Checked 2026-10-13

Dataset	Metric	Score
MTEB multilingual v2	Mean(Task), Multiple	61.36
MTEB code v1	Mean(Task), NDCG@10	78.68

Benchmarks

No submitted benchmark reports yet.