lufficcv/embeddinggemma-2-vision-GGUF

EmbeddingGemma 2: image encoder only, 124 MB

The vision encoder of google/embeddinggemma-2 as an mmproj GGUF for llama.cpp. It's made for on-device photo search, where the audio encoder isn't needed.

It's made from ggml-org/embeddinggemma-2-GGUF at revision bfcd298762cc34d0357ece5ebdd31791a3a374d8:

• mmproj-embeddinggemma-2-Q8_0.gguf with the audio encoder removed (the a.* tensors, the clip.audio.* keys, and clip.has_audio_encoder = false);
• v.position_embd.weight in Q8_0 instead of F32;
• the Q8_0 matrices of the vision blocks requantized to Q5_0 from mmproj-embeddinggemma-2-BF16.gguf;
• everything else (norms, projector) as in the Q8_0 file.

vision-model.py rebuilds it byte for byte: python vision-model.py mmproj-embeddinggemma-2-Q8_0.gguf mmproj-embeddinggemma-2-BF16.gguf out.gguf.

file	bytes	sha256
mmproj-embeddinggemma-2-vision-Q5_0.gguf	124,352,192	93c8369c7883d6f64ef860a52c911835c1c75668647b908b2e04e5f13058bfca

Quality

Text-to-image retrieval on XM3600 (3,600 photos; 7,174 Chinese and 7,200 English captions), image budget 140. The text model is unsloth/embeddinggemma-2-GGUF UD-Q4_K_XL, except in the first row:

setup	zh R@1	zh R@10	en R@1	en R@10
ggml-org Q8_0 text + Q8_0 mmproj (865 MB)	69.6	93.5	59.7	88.8
Q4_K_XL text + this file (300 MB)	69.3	93.2	59.6	88.8

Use it with the same text model it was evaluated with. Image vectors pass through the text model, so they depend on which text model you use.

License

Apache 2.0, the same as EmbeddingGemma 2. See the Gemma 4 license.