Yandex has released the Alice AI Foundation LLM, an open-source large language model trained from scratch. The model is available under the Apache 2.0 license for research and commercial use. It serves as an experimental platform for testing architectural solutions for future unified reasoning models and will underpin the agentic capabilities of Alice AI.
The model utilizes a Mixture of Experts architecture with 80 billion total parameters, of which 3 billion are active during inference. This design allows for high processing speed and efficiency. The developers also improved the training optimizer, which accelerated the training process by approximately two times, and refined the preparation of high-quality training data.
In performance benchmarks, Alice AI Foundation LLM matches Qwen3.5-35B-A3B-Base in math tasks and outperforms Nvidia Nemotron-3-Super-120B-Base in code generation despite having fewer active parameters. It also surpasses larger models like DeepSeek-V4-Flash-Base in factual knowledge tasks and performs on par with the previous closed-source Alice AI LLM, which is significantly larger.
To evaluate performance in Russian, Yandex developed two proprietary benchmarks: WikiWebFacts and HardMultiQA. WikiWebFacts tests knowledge of dates, definitions, and events using encyclopedic data and search queries. HardMultiQA is based on anonymized user requests to Alice AI, focusing on specialized knowledge areas that standard English-language benchmarks cannot objectively assess.
