AliceAI-Foundation-80B-A3B-Base: Achieving 80B-class inference performance with 3B active parameters

It adopts a Mixture-of-Experts structure that activates only 3 billion parameters per token out of a total of 80 billion. By combining a hybrid architecture with KDA layers, it processes long contexts of 262,000 tokens while improving computational efficiency. Through full pre-training, it possesses complex reasoning and tool-use capabilities.

It demonstrates overwhelming performance on Russian factual knowledge benchmarks. It scored 86.5 on WikiWebFacts and 67.9 on HardMultiQA, significantly outperforming competing models such as Qwen3.5 and GLM-4.5-Air. Its mathematical and coding abilities are also top-tier, achieving a high accuracy of 96.7 on AIME 2026 problems.

It can be deployed via the vLLM and Transformers libraries. The flash-linear-attention library is required for GPU environments. Released under the Apache 2.0 license, it is available for commercial use and is suitable for fine-tuning based on Russian-specific datasets.