EVIE 4.5B released

Tencent open-sourced EVIE-4.5B, a compressed sibling to its larger EVIE-8B model, built for visual document retrieval at a fraction of the storage cost. Its embeddings are elastic, meaning a single trained projection can be truncated at runtime to anywhere from 64 to 2048 dimensions depending on how much accuracy a given deployment needs. The bigger change is a training-free token-compression method that shrinks roughly 750 page tokens down to 32 through clustering, which brings a multi-vector retrieval index for a million pages under 4 gigabytes. Despite that compression, the model still leads the ViDoRe V3 benchmark, so the usual tradeoff between index size and retrieval quality mostly disappears here. For teams running retrieval at scale, this is the more practical of the two Tencent releases. It offers nearly the same accuracy as the full-size teacher model at a small fraction of the storage footprint.