> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Yandex releases open weights for its Alice AI Foundation base model
- URL: https://globalfeed.ai/en/yandex-releases-open-weights-for-its-alice-ai-foundation-base-model/
- Published: 2026-09-23T03:27:46.000Z
- Updated: 2026-09-23T14:29:38.000Z
- Description: Yandex has published Alice AI Foundation 80B-A3B Base on Hugging Face under Apache 2.0. The model was trained from scratch, has 80 billion parameters with 3 billion active per token, and covers Russian and English.
- Author: GlobalFeed Editor
- Tags: Yandex, open models, AI models, Hugging Face, x-yandexcom, dil-en, elle

Yandex has published the weights of **Alice AI Foundation 80B-A3B Base** on Hugging Face. The model is released under the Apache 2.0 license, so it can be used in commercial projects, modified and built upon. Yandex says the model was trained from scratch. The family shares its name with Alice, the company's voice assistant.

## What the model is

Alice AI Foundation 80B-A3B Base is a base model built on a mixture-of-experts (MoE) architecture. It has 80 billion parameters in total, but only 3 billion of them are active for each token it generates. The design aims to give the knowledge of a large model at a cost closer to that of a small one. It can read up to 256K tokens of context in one go and was trained for Russian and English.

"Base" means the model has only been pre-trained. It has not been tuned for chat or for following instructions. It is offered as a starting point that developers fine-tune for their own tasks. Yandex says it added complex reasoning and tool-use data to the training corpus.

## What Yandex claims

The benchmarks on Yandex's model card compare it with the base versions of Qwen3.5-35B-A3B, GLM-4.5-Air, Nemotron-3-Super and DeepSeek-V4-Flash. According to Yandex, the model is on par with larger open models in maths and coding and stands out on Russian factual knowledge. It trails some of the compared models on several English knowledge tests.

Alongside the model, Yandex released two new benchmarks for Russian factual knowledge: WikiWebFacts and HardMultiQA.

## How to use it

The model can be downloaded from Hugging Face and run with Transformers or vLLM; the model card has step-by-step commands for both. Because the architecture is new, loading it in Transformers requires allowing Yandex's custom code from the repository to run. For vLLM, Yandex provides a ready Docker image.

For fine-tuning on your own data, the repository includes a LoRA example and a chat template for preparing training data.

## Things to keep in mind

- Although only 3 billion parameters are active per token, all 80 billion still have to fit in memory. Yandex's examples run on several data-centre GPUs; the model is sized for data-centre hardware.
- As a base model, it can give inconsistent results when used directly for chat or instructions. Its main use is fine-tuning.
- It was trained for Russian and English only.
- The model runs Yandex's custom code from the repository. In company environments it is good practice to review that code before running it.
- For a fuller picture, follow evaluations from independent researchers as well.

Sources: [Hugging Face: AliceAI-Foundation-80B-A3B-Base model card](https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base?ref=globalfeed.ai), [WikiWebFacts benchmark](https://huggingface.co/datasets/yandex/WikiWebFacts?ref=globalfeed.ai), [HardMultiQA benchmark](https://huggingface.co/datasets/yandex/HardMultiQA?ref=globalfeed.ai).