Yandex Releases Pretrained Model for Alice's Quick Answers in Search

Yandex has released the open pretrained model AliceAI-T5-35B-A0.6B Base on Hugging Face — a from-scratch encoder-decoder model with sparse MoE following the UL2 recipe, where about 0.6 billion parameters are active per token out of 34.35 billion total — created for Alice AI's quick answers in Search.

Compiled by Sergey KostenchukPublished 2026-09-11Updated 2026-09-18

2026-09-11 Research HuggingFace

🤖 Yandex Releases Pretrained Model for Alice's Quick Answers in Search

The open-weights model AliceAI-T5-35B-A0.6B Base has been released on Hugging Face: an encoder-decoder with sparse MoE following the UL2 recipe, 512 experts per layer with top-8 routing. A total of 34.35 billion parameters, with only ~0.6 billion active per token; context — 128K. Yandex's benchmarks: strong factuality (WikiWebFacts 81.3) and extraction at 32K (Ruler 94.7), but it lags behind Qwen 3.5 35B-A3B Base on MATH 500 and MMLU.

🌍 Ultra-sparse MoE with cheap inference is already in production: the fine-tuned Alice AI Search responds in seconds under hundreds of thousands of requests per minute. A template for search/QA/RAG teams.

👤 Weights and code are open: the model can be downloaded, run via Transformers, and fine-tuned using a ready-made LoRA example — a real reference for open models in Russian.

Source 1: https://huggingface.co/yandex/AliceAI-T5-35B-A0.6B

Source 2: https://habr.com/ru/companies/yandex/articles/1080654/

Sources