Yandex releases AliceAI-Foundation 80B model under Apache 2.0 license
In brief
- Yandex released AliceAI-Foundation-80B-A3B-Base, an 80-billion-parameter LLM on Hugging Face under Apache 2.0 license
- Model features 262,144-token context window and supports unrestricted commercial and non-commercial use
- Weights went live September 18, 2026; full documentation released September 21
Architecture and performance
The model packs 80 billion total parameters spread across 512 routed experts plus one shared expert, organized into 48 layers. Only 3 billion parameters activate at a time, using a Mixture-of-Experts design that reduces computational overhead while maintaining competitive performance. The context window is 262,144 tokens, sufficient for an entire novel, codebase, or extended legal document.
On Yandex's two new Russian-language benchmarks—WikiWebFacts and HardMultiQA—AliceAI-Foundation scored 86.5 and 67.9 respectively. The model matches or beats DeepSeek-V4-Flash-Base in Russian factual knowledge, law, medicine, and coding while using fewer active parameters.
Open release and licensing
The weights are live on Hugging Face, available to anyone who wants to build with them, commercially or otherwise, no strings attached. Apache 2.0 permits commercial use, modification, distribution, and sublicensing without owing Yandex anything beyond attribution. A company can take AliceAI-Foundation-80B-A3B-Base, fine-tune it, and sell the result.
The release as published is a base model only. There is no instruction-tuned or chat-fine-tuned version included out of the box. Developers will need to handle fine-tuning themselves if they want a conversational or task-specific variant.
Future direction
Yandex described AliceAI-Foundation as the base for future work toward a unified reasoning model with stronger agentic capabilities. The open-source release removes barriers for researchers and builders who want to experiment with the architecture, benchmark it against proprietary models, or adapt it for domain-specific tasks.
