Yandex releases AliceAI Foundation 80B-A3B base model under open-source license

Yandex has released AliceAI-Foundation-80B-A3B-Base, a large language model built entirely from scratch and published under the Apache 2.0 license. The weights are live on Hugging Face, available to anyone who wants to build with them, commercially or otherwise, no strings attached.

What the model actually is

The architecture is a Mixture-of-Experts design. The model has 80 billion total parameters spread across 512 routed experts plus one shared expert, organized into 48 layers. For any single token it processes, only 3 billion of those parameters are active.

The context window is 262,144 tokens. For reference, that is long enough to feed the model an entire novel, a full codebase, or an extended legal document without truncating anything.

The release as published is a base model only. There is no instruction-tuned or chat-fine-tuned version included out of the box. Researchers and developers get the raw pre-trained weights, which they can then adapt for their specific use cases.

How it performs against the competition

Yandex published two new Russian-language benchmarks alongside the model: WikiWebFacts and HardMultiQA. AliceAI-Foundation scored 86.5 on WikiWebFacts and 67.9 on HardMultiQA.

The independent comparison that matters more is against DeepSeek-V4-Flash-Base, which the Yandex team cited as a direct competitor. AliceAI-Foundation matches or beats that model in Russian factual knowledge, law, medicine, and coding, according to Yandex’s published evaluations. DeepSeek-V4-Flash-Base activates substantially more parameters per token, which means AliceAI-Foundation is achieving competitive results at lower computational cost.

Math and coding performance also held up against models with significantly larger active parameter counts.

Why the open-source choice is significant

Apache 2.0 allows commercial use, modification, distribution, and sublicensing. A company can take AliceAI-Foundation-80B-A3B-Base, fine-tune it for a specific vertical, wrap a product around it, and sell that product without owing Yandex anything beyond attribution.

Weights were published on September 18, 2026, with the full announcement and documentation following on September 21.

Yandex described AliceAI-Foundation as the base for future work toward a unified reasoning model with stronger agentic capabilities.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.