“Yandex” publish Alice AI Search Pretrain, the artificial intelligence model trained from scratch that forms the basis of the Search Quick Answers service, is now publicly available.
Image source: yandex.ru/company
The compact language model Alice AI Search Pretrain has passed the first stage of training and has a large amount of knowledge about the world. Its compact size allows it to utilize modest computing power, withstand huge loads and process large files.
The model’s hybrid architecture has previously been used mainly in scientific research: it uses encoder-decoder and MoE (Mixture of Experts) techniques. The first one helps her answer concisely and clearly because one part of the model looks for the message and another part makes up the answer. The second technique allows it to process only those fragments of itself that are most relevant to the subject of the request: with a parameter size of 35 billion, there are only 600 million connections to process a token – less than 2% of the total.
Alice AI Search Pretrain, developed by Yandex, outperforms other small language models: Blind test measurements show that its answers are of higher quality than Alibaba Qwen 3.5 2B and 4B and Google T5 Gemma 2 4B-4B Base – its responses are roughly on par with the more computationally demanding Qwen 3.5 35B-A3B. Yandex Search uses an additional trained version of the Alice AI search model – which was prepared taking into account real examples of interactions between users of the search service.
If you find an error, select it with your mouse and press CTRL+ENTER.
