Yandex cuts GPU use severalfold in Alice AI's Research mode without retraining the model
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
Deep research by SI is getting cheaper not only thanks to new models but thanks to smart "packaging" around them, which means it will soon go mainstream.
The team rewrote the scaffolding around the model: processing of search snippets was handed to a trained classifier instead of an LLM, tools are called only if they advance the research, and planning was moved to a lighter model. Over three months the cost of a single query fell 6.6 times; on the FRAMES and SimpleQA benchmarks the mode scores 75 and 81. Individual scaffolding improvements each gave a gain of 5 to 9 percentage points.
Sources 1