Yandex has introduced Alice AI Foundation LLM, an experimental model built on the Mixture of Experts architecture. The model features 80 billion parameters in total, with only 3 billion active parameters utilized during query processing to reduce computational requirements.
The company developed the model entirely in-house without foreign components. It demonstrates strong performance in programming, mathematics, and Russian language tasks, outperforming Nvidia Nemotron-3-Super-120B-Base in coding tests and DeepSeek-V4-Flash-Base in fact-based queries.
To evaluate performance, Yandex released two open-source benchmarks: WikiWebFacts and HardMultiQA. These tools, along with the model and technical documentation, are now available on the Hugging Face platform for public access.
The model serves as a foundation for testing architectural solutions for future agentic functions within the Alice AI assistant. Yandex intends to enable users to delegate specific tasks to the assistant through these upcoming capabilities.
