Octen-WebGR-4B: A 4B Generative Retriever for 200M URLs

Octen has released Octen-WebGR-4B, a 4B generative retriever for navigational web search that maps queries to a table of nearly 196 million real URLs using a two-token semantic identifier. The model stores the web index directly in its parameters, requiring about 10 GB of memory for retrieval instead of a large vector database. While it outperforms dense retrieval on common queries and new ways of asking for familiar pages, it is weaker on rarely visited pages and does not support informational queries. Open article →