Yandex Open-Sources New AI Model Trained from Scratch
Yandex has open-sourced AliceAI-Foundation-80B-A3B-Base, the pretrained version of its new large language model, trained from scratch. It matches larger open-source models in coding and reasoning and requires relatively little inference compute for a model of its size. Developers can use it in both research and commercial projects under the Apache 2.0 license.
The new model is an experimental version that Yandex is using to test architectural decisions for its future unified reasoning model. That reasoning model will power the agentic capabilities of Alice AI, the company's AI assistant, letting users delegate specific actions to it.
Model quality
AliceAI-Foundation-80B-A3B-Base uses a Mixture of Experts (MoE) architecture and has 80 billion parameters, of which only 3 billion are active at any given moment. This significantly reduces the compute required for inference.
The model is built for reasoning and agentic scenarios, which makes it highly effective at complex logic tasks and coding. For example, it solves the vast majority of olympiad-level math problems and shares the lead with Qwen3.5-35B-A3B-Base. On code-writing tests, it outperforms NVIDIA's Nemotron-3-Super-120B-Base while using a quarter of the active parameters. Strong reasoning and coding skills are essential for agents that carry out actions on a user's behalf.
The model is strongest in Russian. In Yandex's evaluations, it outperforms every open base model it was compared with on Russian-language benchmarks for factual knowledge, Russian language, literature, history, law, and medicine. That includes larger models such as DeepSeek-V4-Flash-Base, which has 284 billion parameters and 13 billion active. On the factual benchmarks, the new model also performs at the level of Yandex's previous closed model, Alice AI LLM, which has three times as many parameters and seven times as many active ones.
New benchmarks
Yandex has also released two of its own factual benchmarks, WikiWebFacts and HardMultiQA, which focus on Russian-language contexts. The first consists of short question-and-answer pairs built from online encyclopedias and frequent aggregated search queries. It tests knowledge of dates, definitions, events, and notable people. The second, HardMultiQA, is built on an anonymized stream of queries to Alice AI and covers rarer areas of knowledge, from medicine and law to IT and art. Recalling a single fact is not enough here: the model must choose several correct options from a list, name several facts that fit the question, or find a factual error in a text.
Both benchmarks are public, so users and researchers can reproduce the results and compare models. Along with the task sets, Yandex is publishing reference answers and the full evaluation protocol.
Engineering
During model training, Yandex developers improved the optimizer, the program that controls the training process. They rebuilt it so data transfer between GPUs runs in parallel with computation, making optimization steps roughly twice as fast.
Yandex also streamlined the preparation of high-quality training data. Selecting that data required scoring documents with a resource-intensive model, and running it on the entire corpus would have taken more than 200,000 GPU hours. Documents now pass through a cascade of classifiers before that scoring process. At each stage, a more complex and computationally expensive model works with fewer documents. This workflow cut computation by more than tenfold while keeping about 95% of the useful documents. The engineers also significantly expanded the model's knowledge base in law, medicine, and mathematics.
The new model is available under the Apache 2.0 license, which allows free use for commercial and research purposes. AliceAI-Foundation-80B-A3B-Base has gone through the main training stage, which determines its breadth of knowledge, capabilities, and potential, so developers can use it as a foundation for their own projects.
The model, technical reports, and benchmarks are available on Hugging Face.