Yandex Unveils a New AI Model Trained From Scratch
The multinational IT company Yandex has created its own large language model, Alice AI Foundation LLM. Developers have already released the base version of the AI model for public use. Company representatives say it was trained from scratch, without using foreign components.
The arrival of the new model is good news not only for IT professionals but also for ordinary Russian users. In the future, its technologies are expected to be incorporated into the Alice neural network. This should enable it to handle complex requests to the virtual assistant more effectively while improving its reasoning and ability to complete multistep tasks.
Reads What the Audience Is Asking
The AI model has delivered strong results in programming and reasoning, while its Russian-language response quality has outperformed many global counterparts, particularly when it comes to facts of interest to Russian-speaking users. According to Yandex representatives, Alice AI Foundation LLM outperforms large AI models that are publicly available. At the same time, it does not require large amounts of computing power to process new data. Its free Apache 2.0 license makes it suitable for both research and commercial projects. Yandex’s new IT product is an experimental version. The developer is using it to test architectural approaches for its future Yedinaya Rassuzhdayushchaya Model (Unified Reasoning Model), or ERM. In the longer term, the reasoning model is expected to become the foundation for Alice AI’s agent capabilities, allowing users to delegate specific tasks to it.
80 Billion Parameters, With 3 Billion Active at Once
Yandex’s new large language model, Alice AI Foundation LLM, is built on a Mixture of Experts (MoE) architecture. It contains 80 billion parameters, but only 3 billion are active at any given time. This makes it possible to achieve high operating speed and efficiency. That matters because the more efficient an AI model is, the more broadly it can be used by businesses and research teams that do not have access to large-scale computing resources. Yandex’s new development has several other advantages, including reasoning and agentic capabilities, as well as strong performance on complex logic and programming tasks. For example, it solves olympiad-level problems at a level comparable to Qwen3.5-35B-A3B-Base, while outperforming NVIDIA’s Nemotron-3-Super-120B-Base in code generation. High-quality information analysis and programming are essential to the effective operation of agents that can perform actions on a user’s behalf. Despite its compact size and lower computing requirements, Yandex’s new model outperforms the 284-billion-parameter DeepSeek-V4-Flash-Base on problem-solving tasks. Research has shown that this Russian IT product answers users at the level of Yandex’s previous closed model, Alice AILLM, even though that model had three times as many parameters and seven times as many active parameters.
The developer decided to assess the model’s Russian-language knowledge using new system-testing tools, or benchmarks. The IT team created two proprietary benchmarks, WikiWebFacts and HardMultiQA. The first tests dates, definitions, events, and people. The second covers more specialized areas of knowledge, from medicine and law to IT and the arts.
While developing the new model, Yandex’s IT professionals improved the optimizer, the software that manages the training process. This cut the amount of computation required by tens of times while preserving about 95% of useful documents. Engineers also expanded the AI model’s knowledge base in law, medicine, and mathematics. The new Yandex model has now completed its main training phase and can be used as a foundation for building custom projects. It is available under the Apache 2.0 license and can be used for commercial and research purposes.
Want to Fine-Tune It for Your Own Tasks? Easy!
The arrival of Yandex’s new large language model reflects a broader trend in the IT industry: developers are looking not simply to make AI models larger but to make them more capable while reducing their computing requirements. Russian IT professionals are moving in the same direction as the global market, developing their own digital solutions and investing in their efficiency. According to Yandex representatives, Alice AI Foundation LLM will become a technological foundation for the company’s next generations of products. Other developers can also use the large language model because it can be downloaded and used for commercial purposes. Notably, all information can remain within the user’s own infrastructure, and the AI model can be fine-tuned for a specific industry.
At this point, Yandex’s new model is likely to be most useful to specialized professionals, but its technologies are expected to expand the capabilities of the Alice neural network and make it more effective. Over the longer term, this new model could serve as a foundation for enterprise assistants, automated support, and RAG systems that combine large language models with external databases. Examples already exist: YandexGPT 5 is used in Yandex Cloud to build AI assistants.
The release of the new large language model is an important step in the development of Russia’s AI industry. Alice AI Foundation LLM could potentially become an export product for countries where Russian is used and Russian-speaking communities live. However, it is too early to say so, since the model still needs to be developed from a base version into a finished product.
