Yandex has released AliceAI-Foundation-80B-A3B-Base, an 80 billion-parameter language model, under the Apache 2.0 license. The model is available on Hugging Face for commercial and academic use without licensing fees.
The model utilizes a Mixture-of-Experts architecture, activating 3 billion parameters per token. It selects from 512 experts, with 10 active per token plus one shared expert. The system features 48 layers and supports a 262,144-token context window.
Trained on 18 trillion tokens, the model outperformed DeepSeek-V4-Flash-Base on the new WikiWebFacts and HardMultiQA benchmarks. It also competes with NVIDIA's Nemotron-3-Super-120B in coding tasks.
Yandex positions the model as a sovereign AI solution, emphasizing its accuracy in Russian legal and medical contexts. The lower active parameter count is intended to reduce hardware requirements for on-premise deployment.
