Yandex releases 80B MoE base model under Apache 2.0

Yandex has published AliceAI-Foundation-80B-A3B-Base, an 80-billion-parameter mixture-of-experts (MoE) language model released under the permissive Apache 2.0 license. The model is available now on Hugging Face.

Despite its 80B total size, the model activates only about 3B parameters per token — the "A3B" in its name. That sparse design is the core appeal of MoE architectures: they aim to deliver the knowledge capacity of a large model while keeping inference costs closer to those of a much smaller dense one. It targets text in Russian and English.

Why it matters

This is a base model rather than an instruction-tuned chat variant, which makes it a foundation for developers who want to fine-tune for their own tasks. A few reasons it stands out:

• The Apache 2.0 license permits commercial use and modification with minimal restrictions.
• Strong Russian-language capability is comparatively rare among openly licensed frontier-scale models.
• The MoE design lowers the practical compute barrier for teams that want large-model quality.

As an initial release, the model arrives without a stated context length or published benchmark results in the record, so real-world evaluation will depend on the community putting it through its paces. Still, the combination of scale, an efficient architecture, and a genuinely open license makes AliceAI-Foundation a notable entry from a major player outside the usual US and Chinese labs.