EmbeddingGemma 2: Google's 740M multimodal embedding model, Apache 2.0

Google's Gemma release notes list EmbeddingGemma 2 on October 6, 2026, "in 740M parameter size." The release notes give no other specs. The details below come from Google's model card and the matching Hugging Face card (`google/embeddinggemma-2`), both updated October 6, 2026, and from Google's launch post of the same day. - **Size:** 740M parameters in total: a 270M text model plus separate vision (170M) and audio (300M) encoders. The card lists a text-only load at 270M. - **Inputs:** text (including code), images, video and audio, mapped into one embedding space. - **Output:** 768-dimension embeddings, with Matryoshka (MRL) truncation to 512, 256 or 128. The card says to re-normalize vectors after truncating. - **Context:** one 8,192-token window shared by all modalities. - **Languages:** Google says it understands 100+ languages. - **Prompts:** the card says the model is trained with short task prefixes on text inputs. Leaving them out still works but lowers precision. - **Precision:** bfloat16 or float32, not float16. The card recommends bfloat16 on hardware with native support and float32 elsewhere, including most CPUs. - **Formats:** Google's Hugging Face repo holds safetensors weights, and the card's code uses sentence-transformers. The launch post lists transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LM Studio for serving it. It links to a GGUF conversion by ggml-org and to an Ollama library page. - **Target devices:** Google says it is designed for consumer hardware such as mobile devices and laptops. - **Licence:** Apache 2.0, per the launch post and the Hugging Face tag. The card's licence link leads to the Apache License 2.0 text. The card also says deployments must follow the Gemma Prohibited Use Policy. We have not benchmarked EmbeddingGemma 2.