EmbeddingGemma 2

Commenters largely welcomed EmbeddingGemma 2’s Apache 2.0 license and open weights, with several saying that having a hosted option plus fallback to runnable weights is essential for embedding workloads that store millions of vectors (simonw). Many praised its multimodal capability (text, vision, audio), the modest parameter counts (reported as 740M total with 270M for text) and the ability to enable only needed modalities (sourcecodeplz, djoldman). Enthusiasts said the model fills a gap for moderate-size embeddings and could power local embedding tools and multimodal search (minimaxir, nowittyusername, alberto467), while others noted it could be useful for "Jev"-like tasks on-device (Nautman) and commented on how image/audio embeddings can carry both content and style signals (alberto467, onlyrealcuzzo).

Opinion divides on comparisons and technical tradeoffs. Some urged benchmark comparisons with other offerings - VoyageAI, Qwen, and Google’s Siglip2 - with one commenter reporting VoyageAI as better in their experience (dcl, brokensegue, nostrebored). Aabhay pointed out a methodological difference: Gemma 2 was trained with MRL rather than MatFormers, meaning you can’t shrink model weights alongside lower-dimensional embeddings, a limitation for on-device sizing. Others questioned how well multimodal proximity and alignment will work in practice and flagged reliance on providers to cover re-embedding costs as an unresolved risk (simonw).