riso348/embeddinggemma-2-text

EmbeddingGemma 2 — text only

The text tower of google/embeddinggemma-2 without the vision and audio encoders: 542 MB instead of 1.5 GB, same text embeddings.

• What changed: only the language_model.* tensors (413, bfloat16) are kept in model.safetensors. Nothing was retrained, merged or re-quantized. Config, tokenizer and the sentence-transformers modules (mean pooling, normalization, prompts) are copied unchanged.
• Embeddings: identical to the original for text input (cosine 1.00000 in float32; ≥ 0.99994 in bfloat16 on Apple GPUs).
• sentence-transformers: loads as usual and gives the same text embeddings (cosine 1.0). It warns that the vision and audio weights are missing and initializes those towers randomly; they are never used for text.
• Use: text search on phones and laptops, where downloading the image and audio parts is wasted. Prompts as in the original: query task: search result | query: , document title: none | text: .
• Made for: on-device mail search in Enveloo (Rust, candle). The extraction script is scripts/extract_embeddinggemma_text.py in the Enveloo repository.

License

Apache 2.0, the same as the original model (Gemma 4 license). This is a modified version of google/embeddinggemma-2: the vision and audio weights were removed. All credit for the model goes to Google DeepMind.