Simon Willison Highlights Why Open Models Like EmbeddingGemma 2 Prevent Vendor Lock-In
Google released EmbeddingGemma 2 under the open-source Apache 2.0 license, offering a natively multimodal open embedding model. Prominent developer Simon Willison highlighted that open-license embedding models are essential to protect developers from costly vector re-computation if hosted services deprecate proprietary models. Applications using embeddings often store millions of vectors for semantic search and retrieval; if a vendor deprecates a proprietary model, re-embedding all stored data requires substantial financial and compute costs. Open-weights models under permissive licenses like Apache 2.0 allow developers to use cloud APIs with peace of mind, knowing they can self-host or migrate the exact model if the original provider shuts it down. EmbeddingGemma 2 natively maps text, images, audio, and video into a unified vector space, supporting multimodal search applications. Additionally, it uses Matryoshka Representation Learning (MRL), allowing its 768-dimensional native vectors to be safely truncated down to 512, 256, or 128 dimensions to lower storage demands.
## BACKGROUND
An embedding model converts unstructured input into dense numerical vectors such that semantically similar items sit close together in vector space. Because vector spaces are model-specific, changing or upgrading an embedding model breaks compatibility with previously generated vectors, requiring every stored document to be re-indexed.