Dojo-Embedding v1.0 (740M)

The on-device multimodal embedding model used by Dojo for local search over text, images and audio. Text, images and audio map into one 768-dimensional vector space.

Dojo-Embedding v1.0 is google/embeddinggemma-2, unmodified, quantized to GGUF, with the multimodal projector split into separate vision and audio files and renamed for Dojo. No fine-tuning. All credit for the model belongs to Google DeepMind.

Files

The text model works on its own; the encoders are optional and load only when images or audio are embedded.

File	Parameters	Quantization	Purpose
`Dojo-Embedding-v1.0-270M-Q8_0.gguf`	270M	Q8_0	Text backbone and embedder
`mmproj-Dojo-Embedding-v1.0-vision-170M-Q8_0.gguf`	170M	Q8_0	Vision encoder
`mmproj-Dojo-Embedding-v1.0-audio-300M-Q8_0.gguf`	300M	Q8_0	Audio encoder

License

Apache License 2.0, the same as the upstream EmbeddingGemma 2 model. See the upstream model card for intended use and limitations.

Downloads last month: 82

GGUF

Model size

0.3B params

Architecture

gemma-embedding2

Hardware compatibility

8-bit

Model tree for DojoWorkspace/Dojo-Embedding-v1.0-740M-gguf

Base model

google/embeddinggemma-2

Quantized

(116)

this model