Yandex open-sources Alice AI Search Pretrain, the base model behind AI answers in Search
Yandex
Yandex published the pretrain (base) checkpoint of Alice AI Search on Hugging Face — a compact from-scratch model with a hybrid encoder-decoder plus Mixture-of-Experts architecture, a combination Yandex says previously existed mostly in research. The fine-tuned version has been generating the AI answers shown under the Yandex Search box since July. In blind tests on Russian it beats compact models like Qwen 3.5 2B/4B and T5 Gemma 2 4B Base and matches Qwen 3.5 35B-A3B at a fraction of the compute; Yandex released the pretrain stage rather than the instruction-tuned production model.
Why it matters
A rare case of a lab open-sourcing the pretrain base behind a live production search-AI stack, and one of few public encoder-decoder + MoE models anywhere — a notable base for the Russian-language open-model ecosystem.
Importance: 3/5
Notable open-weights release with official Habr post, HF weights and CNews coverage