Yandex has released the AliceAI-Foundation-80B-A3B-Base language model under the Apache 2.0 license. The model was trained from scratch without using weights from third-party open-source models. Development took approximately six months, involving a complete overhaul of the training corpus, architectural adjustments, and hyperparameter tuning.
The model features 80 billion parameters and is designed to support the agentic capabilities of the Alice AI assistant. In pre-training benchmarks, it outperforms larger models like DeepSeek-V4-Flash-Base and Nemotron-3-Super in areas including mathematics, coding, and factual knowledge. It also surpasses the previous closed-source Alice AI LLM 235B model while having significantly fewer active parameters.
Yandex published a technical report detailing the training process, including data preparation pipelines, scaling laws, and ablation studies. The team conducted experiments using 2 trillion tokens per training run. The company also released two new open-source factual benchmarks, WikiWebFacts and HardMultiQA, to assist the community in evaluating model performance.
The model demonstrates strong performance in Russian-language tasks, including culture, history, and medicine. It leads in benchmarks such as IMO AnswerBench, HMMT, and LiveCodeBench. The final checkpoint of the model is available on Hugging Face, intended for use as a foundation for post-training and product development.
