Yandex has released the AliceAI-Foundation-80B-A3B-Base language model, which is now available on Hugging Face. The model is published under the Apache 2.0 license, allowing for commercial use, modification, and distribution without additional fees.
The model utilizes a Mixture-of-Experts architecture featuring 80 billion total parameters across 512 routed experts and one shared expert, organized into 48 layers. During processing, only 3 billion parameters are active per token, which reduces computational costs.
The architecture supports a context window of 262,144 tokens. Yandex released the base model weights on September 18, 2026, with full documentation following on September 21, 2026. No instruction-tuned or chat-fine-tuned versions are included in this release.
Yandex introduced two benchmarks, WikiWebFacts and HardMultiQA, where the model scored 86.5 and 67.9 respectively. The company claims the model matches or outperforms DeepSeek-V4-Flash-Base in Russian-language factual knowledge, law, medicine, and coding tasks.
The company intends to use AliceAI-Foundation as a base for developing future unified reasoning models with enhanced agentic capabilities.
