Octen-WebGR, a 4B model that memorized 200M URLs (open weights, live demo) (huggingface.co)

Octen has unveiled Octen-WebGR, a novel generative retrieval model with 4 billion parameters that effectively memorizes 195.9 million URLs. This model, designed specifically for navigational web search, generates a unique two-token semantic identifier (SID) for any user query, enabling efficient retrieval directly from its parameters. Unlike traditional models that require substantial storage for vectors, Octen-WebGR operates with only one GPU and about 10 GB of memory, showcasing a significant step forward in resource efficiency. The model is commercially accessible via the Pulse Search API, which boasts a quick response latency of around 10 milliseconds, with usage starting at just $0.1 per 1,000 calls. The development of Octen-WebGR is significant for the AI/ML community as it redefines generative retrieval capabilities, particularly for handling large-scale web data. Its training on query-URL click pair interactions up to August 2026 and subsequent performance metrics indicates strong effectiveness in ranking results based on user behavior. With a hit rate of 59.6% at the top position for relevant queries, this model can outperform traditional dense retrieval approaches, especially for frequently accessed web pages. Furthermore, by limiting the output to a predefined table of URLs, it enhances the reliability and safety of the search process, ensuring that the returned links are appropriate and free from harmful content.