EmbeddingGemma 2, text only, 8-bit MLX
google/embeddinggemma-2 (revision
914f7f89142e33e77833254d9c9b90c3cef7303b) with its image and audio encoders removed, for text search on Apple
silicon. 320 MB in all, against about 1.27 GB for the 8-bit conversion that keeps
the encoders.
- Text tower only (271M of the model's 744M parameters).
- Matrices quantized affine, 8 bits, group 64; scales, biases, norms and scalars in bfloat16 (never float16: the model overflows in float16).
- 768-dimensional output with Matryoshka prefixes 512, 256 and 128; mean pooling,
L2 normalization. Use the source model's prompts, for example
task: search result | query: …andtitle: none | text: …. - The tokenizer and the sentence-transformers configs are Google's files, byte for byte.
Converted by RuniT-ChatFramework's Scripts/convert-embeddinggemma2-text.py. Its
vectors match the FP32 reference cosines of Google's checkpoint within 0.002 on
multilingual query and document pairs, the same as the 8-bit build with encoders.
License: Apache 2.0, as the source model (https://ai.google.dev/gemma/docs/gemma_4_license).