Alice AI Foundation LLM: Yandex's open AI model outperforms larger neural networks

⚡ The gist in 5 seconds - Yandex has released the pretrain version of its large language model Alice AI Foundation LLM, trained entirely from scratch. - The model is available for research and commercial projects under the Apache 2.0 license. - This is an experimental version for testing architectural solutions ahead of a future unified reasoning model, not a final product. ### 🔍 What was found On September 21, 2026, Yandex unveiled and open-sourced the pretrain version of the Alice AI Foundation LLM model. It is built on a Mixture of Experts architecture: 80 billion parameters in total, with only 3 billion active at any given moment. This is exactly what ensures high request processing speed and efficient inference without massive computing power. The developers built into the model capabilities for reasoning, programming, and working in agentic scenarios. According to the company, in olympiad-level math tasks Alice AI Foundation LLM shares the lead with Qwen3.5-35B-A3B-Base. In code generation benchmarks, it outperformed Nvidia Nemotron-3-Super-120B-Base while having four times fewer active parameters. In broad factual knowledge tasks, the model beats DeepSeek-V4-Flash-Base (284 billion parameters, 13 billion active). Moreover, its responses are on par with the previous closed Alice AI LLM, which has three times more parameters and seven times more active ones. ### 💡 Why it matters The practical benefit of the release is already clear: the open Apache 2.0 license allows the model to be used both in research and in commercial projects. A separate emphasis is placed on knowledge evaluation for the Russian-speaking audience — Yandex created its own benchmarks, WikiWebFacts and HardMultiQA. The first tests knowledge of dates, definitions, events, and personalities based on online encyclopedias and search queries, while the second is built on anonymized queries to Alice AI and covers rare areas of knowledge. English-language analogs, the developers note, do not allow an objective evaluation of Russian-language models. ### 🧩 Context Alice AI Foundation LLM is an experimental model that Yandex uses to validate architectural solutions before building a unified reasoning model. It is this model that will form the basis of Alice AI's agentic capabilities: users will be able to delegate various tasks to the system. During training, the company improved the optimizer responsible for managing the training process — this process was sped up roughly twofold. The preparation of high-quality training data was also optimized.

🔗 Read on 3dnews.ru