Yandex has released AliceAI-Foundation-80B-A3B-Base, an open-source language model trained from scratch under the Apache 2.0 license. The model utilizes a Mixture-of-Experts (MoE) architecture, featuring 80 billion total parameters with 3 billion active parameters during inference.
In performance benchmarks, the model achieved 50.5 points on LiveCodeBench v5-6, outperforming DeepSeek-V4-Flash-Base at 38.1 points. On the HMMT Feb 2026 math test, it scored 96.9 points. The model also outperformed Yandex's previous 235B parameter model in 75% of internal tests.
The training process involved a four-stage pipeline, including 17.5 trillion tokens for pre-training, context expansion to 256K, and reasoning training. Developers utilized a 27 billion token mathematical corpus named MaRS, filtered from internet documents with a 1 to 140,000 conversion rate.
Yandex has published a technical report alongside the model weights, detailing data preparation and training stabilization. The release includes two new Russian-language benchmarks, WikiWebFacts and HardMultiQA, which are now available for public use on the Hugging Face platform.
