WeMM Embedding 2B vs WeMM Embedding 9B

WeMM Embedding 2B and 9B encode supported text, images, video and visual documents into a shared retrieval space. They return embeddings rather than chat answers. The smaller model uses 2048-dimensional vectors and the larger model 4096-dimensional vectors, so switching between them affects both encoder deployment and index storage.

Build separate indexes over the same collection with each model’s documented preprocessing and query or document instructions. Use labelled queries that cover paraphrases, visual details and confusingly similar pages. Measure recall and ranking quality, then inspect false matches and missed evidence. Keep downstream answer generation out of the first retrieval comparison so fluent prose does not hide a weak result set.

Record indexing time, vector storage and search latency at the collection size your application expects. The 9B encoder’s larger vectors are not interchangeable with a 2B index; a deployment change requires rebuilding or explicitly versioning the stored representations.