Yandex Releases AliceAI-Foundation-80B-A3B, an Open-Weights 80B Base Model
Tech company Yandex has open-sourced AliceAI-Foundation-80B-A3B-Base under the Apache 2.0 license on Hugging Face. Trained from scratch, this 80-billion parameter base model uses a custom Mixture-of-Experts (MoE) architecture with only ~3 billion active parameters per token. This release provides the open-weights ecosystem with a major new model family that does not rely on standard Llama architectures. Its high parameter count combined with low active parameter usage allows for efficient inference while offering strong competitive performance for non-English and multilingual tasks. The model features a 256K context window and is released as a pure base model without post-training or instruction tuning. Because it relies on a custom architecture, popular inference frameworks like llama.cpp do not support it out of the box yet.
## BACKGROUND
In Mixture-of-Experts (MoE) architectures, the total parameters represent the model's overall capacity, while active parameters (indicated by 'A3B') refer to the subset routed and computed during inference for a single token. Base models are pretrained on massive text datasets to predict next tokens and typically require alignment or post-training before being used as chat assistants.