What it does
The paper covers small answer sets as well as retrieval among millions of candidates, with images and video included. Putting options in the request helps on some tasks, but reasoning is still a bottleneck. Its text-only JevBench results sit slightly below Open-Jev-27B. No released MetaEncoder weights are identified.