Nemosine/embeddinggemma-2-text-coreml

EmbeddingGemma 2 (text) for Core ML and the Apple Neural Engine

A Core ML conversion of the text model of google/embeddinggemma-2 (revision 914f7f89142e33e77833254d9c9b90c3cef7303b), made by Nemosine for on-device search on iOS 18+ and macOS 15+. Converted with nemos/tool/coreml_embedding/convert_gemma2.py.

• Text only (no vision or audio encoders), 768-dimensional output with Matryoshka truncation to 512/256/128 as in the original.
• Float16 weights and compute, no quantization or pruning.
• One multifunction package: s128, s256, s512, s1024, s2048 (fixed sequence lengths sharing one copy of the weights).
• Inputs: input_ids (int32, BOS … EOS, zero-padded) and attention_mask (float32, 1 for tokens, 0 for padding). Output: embedding, a mean-pooled, L2-normalized 768-d vector.
• Prompts are the original model's: task: search result | query: {query} and title: {title | none} | text: {text}.
• Agreement with the original float32 model: see conversion_report.json.

License and modifications

EmbeddingGemma 2 is developed by Google DeepMind and released under the Apache License 2.0 (see LICENSE). This repository contains a modified version; NOTICE describes the modifications. Use is subject to the Gemma Prohibited Use Policy. Google DeepMind and Gemma are named only to identify the origin of the work; this conversion is not endorsed by them.