Yandex company propose And disclosed a pre-trained version of the large-scale Alice AI Foundation LLM language model, which was completely trained from scratch. The solution is said to demonstrate high results in programming and reasoning and outperform many foreign counterparts in terms of the quality of Russian-language answers, especially in factual knowledge tasks that are of interest to Russian-speaking users.
Image source: yandex.ru
According to Yandex, Alice AI Foundation LLM outperforms larger open AI models and does not require large amounts of computing power for inference. Developers can use neural networks while conducting research and in commercial projects under the Apache 2.0 license. The proposed AI model is experimental and used to test architectural solutions for Yandex’s future unified inference model. It will be the basis for agency capabilities «Alice Ai»so users will be able to assign the system to perform different tasks.
Alice AI Foundation LLM operates on an expert hybrid architecture and contains 80 billion parameters, 3 billion of which are currently active. Therefore, high request processing speed and efficiency are achieved. Developers have built into the model the ability to reason and work in agent-based scenarios. Therefore, the system exhibits high efficiency in complex logic tasks and programming.
Alice AI Foundation LLM shares the lead with Qwen3.5-35B-A3B-Base when it comes to solving Math Olympiad problems. In tests of generating code, the product outperformed the Nvidia Nemotron-3-Super-120B-Base neural network, despite having four times fewer active parameters. Despite its compact size, the new AI model outperforms larger open models, including DeepSeek-V4-Flash-Base with 284 billion parameters and 13 billion active parameters, on tasks that require extensive factual knowledge. Benchmark testing shows that Alice AI Foundation LLM is as responsive as the previous closed Alice AI LLM model, which has three times more parameters and seven times more active parameters.
To assess knowledge relevant to Russian-speaking audiences, the developers created two of their own benchmarks – WikiWebFacts and HardMultiQA. The first consists of short question-short answer pairs, suitable for testing knowledge of dates, definitions, events, and personalities based on information from online encyclopedias and frequency-aggregated search queries. The second benchmark is built on an anonymous request stream «Alice Ai »and covers rarer areas of knowledge. Both benchmarks were created to evaluate the knowledge of the Russian model, as the English model does not allow this to be done objectively.
During the Alice AI Foundation LLM training process, developers improved the work of the optimizer (the program that manages the learning process). Therefore, the optimization process is roughly doubled. In addition to this, the developers have also optimized the preparation of high-quality training materials.
If you find an error, select it with your mouse and press CTRL+ENTER.
