Yandex open-sources Alice AI Foundation LLM

Releasing an 80B/3B MoE under Apache 2.0 demonstrates that high-capacity reasoning models can remain accessible and economical to run, establishing an open foundation for practical agentic systems.

• –High inference efficiency: Activating only 3 billion parameters per token lowers hardware requirements and latency while preserving the deep knowledge retention of an 80B model.
• –Tuned for agentic workflows: The model's focus on logic, math, and code synthesis directly targets the core capabilities needed for autonomous AI agents to execute multi-step user tasks.
• –Non-English factual evaluation: Open-sourcing WikiWebFacts and HardMultiQA provides standardized benchmarks and ground-truth protocols for multi-hop Russian factual verification.
• –Training pipeline optimizations: Overlapping GPU communications with compute doubled optimizer step speed, while cascading classifiers cut document filtering compute by orders of magnitude.