Yandex opens weights of AliceAI-Foundation-80B-A3B-Base pretrain

Yandex has released on Hugging Face under the Apache 2.0 license a base model trained from scratch, AliceAI-Foundation-80B-A3B-Base, with 80 billion parameters (3 billion active) and a context of up to 262,144 tokens.

Compiled by Sergey KostenchukPublished 2026-09-21Updated 2026-09-21

2026-09-21 Research HuggingFace

🤖 Yandex opens weights of AliceAI-Foundation-80B-A3B-Base pretrain

Under the Apache 2.0 license, a base model trained from scratch has been published on Hugging Face: 80 billion parameters (3 billion active), MoE with 512 experts, 262,144-token context. According to Yandex, it outperforms larger open pretrains DeepSeek-V4-Flash-Base and GLM-4.5-Air-Base on Russian facts, math, and code.

🌍 Yandex's first pretrain, fully trained from scratch, competes with open models from Chinese labs with fewer active parameters. Apache 2.0 allows commercial use; comparisons with competitors are currently proprietary.

👤 The weights are already available to run: in transformers (reference 5.16.1, flash-linear-attention 0.5.0 required for KDA layers) or in vLLM via a Docker image with tensor-parallel on 4 GPUs. However, this is a base model without chat instructions — for dialogue, your own SFT/RL is needed.

Source 1: https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base

Sources