Yandex trains Alice AI Foundation LLM from scratch with 80 billion parameters

Yandex has presented a new large language model, Alice AI Foundation LLM. It was trained from scratch, without any third-party foreign components, and the company calls it fully sovereign; a Yandex representative told Vedomosti about the development.

The model has 80 billion parameters, of which 3 billion are active, and it was trained on 18 trillion tokens. The weights are open: the license allows the model to be freely run and fine-tuned, including in commercial and research projects. The idea is that it will become the foundation for a future reasoning model and for the agentic capabilities of Alice AI.

The sovereignty requirements are set by the law on artificial intelligence (No. 243-FZ), which came into force on September 1. A sovereign model must be built on entirely domestic code, architecture and infrastructure, whereas a national one permits foreign components, including open-source solutions. The status will be assigned by the government; the decree is still being drafted in dialogue with the industry.

Sber reacted with restraint: the level of Russian engineering teams remains high, but there are already plenty of compact language models of up to 100 billion parameters, and the country's place in the global race is determined by flagships with hundreds of billions and trillions of parameters. In July, German Gref claimed that only the bank has a fully domestic large language model, while Yandex merely fine-tunes Chinese Qwen neural networks; the company denied this.

Market participants call the step significant: open licenses remove legal barriers, and Sber and Yandex are forming two independent centers of competence in foundation models. At the same time, Alice AI Foundation is still a base model rather than a finished assistant: it can become a reasoning one only after fine-tuning.

The status of a sovereign model will only begin to be assigned under the law from March 1, 2027, and until then a public claim of sovereignty remains a marketing characteristic rather than a legally established fact. Experts add that training from scratch is not enough — reproducibility of the entire cycle will have to be confirmed, from data and architecture to training, post-training and inference, and that such a path by itself does not guarantee safety. Yandex is not disclosing the cost of development.

Image: WikiFido CC BY-SA 4.0 · License

Translation editor: Салтанов А.