Yandex has developed a large language model, Alice AI Foundation LLM, which the company trained from scratch without third-party foreign components. The model has 80 billion parameters, with about 3 billion active simultaneously. For training, 18 trillion tokens were used — fragments of text on which the neural network learned to understand and generate information.
The company also opened the model's weights. This allows developers to run it independently and fine-tune it for their tasks. The license permits the use of Alice AI Foundation in commercial and research projects. Yandex expects to use the model as a basis for a future system with reasoning capabilities and agent functions for Alice AI.
However, it is too early to call Alice AI Foundation an officially recognized sovereign model. Yandex positions the development as fully sovereign, but the law establishes a separate procedure for confirming compliance with the requirements for such models. The main norms related to this regulation will come into force on March 1, 2027.
According to the law, a sovereign large fundamental model must be developed by a Russian legal entity, and its life cycle must be technically reproducible. User responses and data storage must be provided by Russian data centers, and the model itself must pass the established compliance confirmation. For a national model, the requirements are softer — the law allows the use of components created both in Russia and abroad, if they are distributed under an open license.
In July, Sber's CEO German Gref said that only the bank had a fully domestic large language model, and Yandex was allegedly engaged in fine-tuning Chinese Qwen neural networks from Alibaba. Yandex denied this. Now the company has presented its own model, trained from scratch.
A Sber representative called the emergence of Alice AI Foundation evidence of the high level of Russian engineering teams. However, he noted that there are already many models with up to 100 billion parameters, and the country's position in the global race, in his opinion, is primarily determined by flagship models with hundreds of billions and trillions of parameters.
Some experts consider the new model part of the competition between Yandex and Sber. Mikhail Koroteev, head of the AI department at the Financial University, called it a response to the ongoing discussion and an attempt by Yandex to close the gap with Sberbank. After post-training, in his opinion, the model will be able to compete with GigaChat 3.5 Reasoning.
Yandex does not disclose how much the development of Alice AI Foundation cost. Expert estimates vary widely. An interlocutor of Vedomosti at a large cloud provider estimated the main training phase at 150–450 million rubles. Valentin Malykh, head of the large language model development center, through an MWS AI representative, named a sum of about 4 billion rubles for renting Nvidia DGX systems. Alexander Gromov, business development director at Data Light, estimated computing costs, including experiments and failed launches, at approximately 0.5–1 billion rubles.
Read more on the topic:
- GigaChat can now change its mind: Sber added a reasoning mode to the neural network
- "Yandex" is developing the first reasoning neural network in Russia
- AI Law Comes into Force: Sovereign and National Models Defined