Octen-WebGR Indexes 201M Web Pages in a 4B Model's Parameters, Beats Dense Retrieval and Frontier LLMs at Navigational Search
_reachsumit · x · 2026-10-28
Octen-WebGR is a generative retrieval framework that stores web pages in a language model's parameters: given a query, the model generates a two-token semantic identifier constrained to a table of real URLs. The 4B variant indexes 201 million pages and was evaluated on real navigational traffic from the month after training.
- On unseen queries for pages with substantial interaction history, it ranks the correct page first 30.9% of the time vs 17.2% for a dense retriever trained on the same clicks (21.4% with a click prior); dense retrieval stays ahead on rarely visited pages, so the two generalize on complementary sides.
- Over all navigational traffic it reaches 59.6% top-1 accuracy vs 46.5% for a click-enhanced dense retriever, while frontier LLMs answering from memory manage only 37.5% at 40x+ the per-query cost.
- The 4B model runs on a single GPU with 10 GB of memory, versus 412 GB for exact dense retrieval.
More from Infra
- LiquidAI Ships 2B Omni Open Model With Day-One llama.cpp Support — ilintar · 2026-10-28
- Mojo/MAX Hits Phones: 4-Day Project Matches llama.cpp Vulkan Inference — Cane_P · 2026-10-28
- EmbeddingGemma 2 runs real-time embeddings on a MacBook CPU alone: 20ms text, 100ms images — generativist · 2026-10-28
- Statistically lossless Qwen3.8-27B GGUF quantizations land on Hugging Face, analysis paywalled — Dany0 · 2026-10-28
- Nebius benchmark: 8 small GPU replicas yield 42.5% more Cosmos video per hour — shashib · 2026-10-28
- Data centers become the new front line as AI, cyber, physical and OT threats converge — ChuckDBrooks · 2026-10-28