Yandex has released the weights of its new language model, AliceAI-Foundation-80B-A3B-Base, under the Apache 2.0 license. The key engineering detail of this release lies in its methodology: the model underwent a full training cycle from scratch, without borrowing external weights or compiling ready-made open-source components.
The development team views this launch as an intermediate step toward creating a unified reasoning model (URM) designed to form the foundation of future enterprise agent systems.
Fewer parameters alongside metric growth
In pre-training evaluations, AliceAI-Foundation-80B-A3B-Base outperforms comparable class models across numerous tasks, including expert queries, mathematics, and code generation. On complex benchmarks such as IMO AnswerBench and LiveCodeBench, the model maintains a leadership position among open pre-trained models.
«It surpasses our previous closed Alice AI LLM 235B in factual knowledge, mathematics, programming, and long-context handling—while using nearly three times fewer total parameters and approximately seven times fewer active ones.»
This discrepancy in architectural efficiency clearly demonstrates how data curation quality and hyperparameter selection have advanced compared to the previous generation of the proprietary lineup.
Experiments and open datasets
The entire path to releasing AliceAI-Foundation-80B-A3B-Base took about six months. To isolate the impact of architectural modifications and the updated corpus, the developers conducted a series of training runs from scratch of 2 trillion tokens each, building on accumulated engineering expertise.
Alongside the model, Yandex open-sourced two specialized datasets—WikiWebFacts and HardMultiQA, with an emphasis on Russian-language context—along with supporting evaluation protocols. In comparative tests, the model ranks first on most factual knowledge and expert skill benchmarks, while delivering top-tier results in its category on MATH-500 and LiveCodeBench.
This release appears to be a pragmatic maneuver: the corporation is not merely sharing weights but setting a new transparency standard for a market weary of taking closed claims at face value. Let us see how competitors respond to this openness.