Yandex's 80B model is available for download: weights released under Apache-2.0

Yandex built Alice AI Foundation 80B-A3B from scratch and trained it on 18T tokens. For each token, only 3B of its 80B parameters are active, so the giant runs at the cost of a smaller model. Its context window is 262K tokens, or hundreds of pages of text in a single pass.

**Russian knowledge leads by a wide margin.** On the WikiWebFacts test, the model scores 86,5 versus 62,4 for Qwen3.5-35B. On the math benchmark MATH-500, the score is 91,1 versus 81,9. The figures come from the model card and were measured by Yandex itself.

For code, the picture is different. CodeForces CPP: 68,9 versus 73,7 for Qwen3.5-35B. This is the only category in the table where Yandex trails.

Three launch paths are described. With vLLM, it is `pip install vllm`, followed by `vllm serve "yandex/AliceAI-Foundation-80B-A3B-Base"`; recipes for SGLang and transformers are available alongside it. The environment needs exact versions: transformers 5.16.1 and flash-linear-attention 0.5.0. The LoRA fine-tuning example is designed for four 80 GB GPUs.

The release is at the pretraining stage, without post-training or alignment. You cannot chat with it as you would with Alice AI, but you can fine-tune it for your own task from day one.

On this foundation, Yandex promises a reasoning model and agent capabilities for Alice.