Yandex open-sources AliceAI-Foundation-80B-A3B-Base, an 80B/3B-active MoE trained from scratch (Apache 2.0)

2026-09-21★★after cutoffopen-sourceYandexconfidence: high

On Sept 21, 2026 Yandex released AliceAI-Foundation-80B-A3B-Base on Hugging Face under Apache 2.0: a pretrained (not instruction-tuned) mixture-of-experts model with 80B total and 3B active parameters, 262K context, trained from scratch. Yandex calls it an experimental testbed for the architecture of its upcoming unified reasoning model for the Alice AI assistant, and claims the best Russian-language results among open models and a win over Nemotron-3-Super-120B-Base.

Key facts

80B total / 3B active parameters; 48 layers, hybrid KDA + gated-attention layers; 512 experts (top-10 + 1 shared); 262,144-token context; vocabulary 129,024 (model card)
License: Apache 2.0; base model only, Russian and English
Selected model-card scores: MATH-500 91.1%, MMLU-Pro (CoT) 66.8%, EGE CoT 90.5%, FinQA 128k 74.1%, WikiWebFacts 86.5%, HardMultiQA 67.9%
Yandex claims: outperforms NVIDIA Nemotron-3-Super-120B-Base with about a quarter of its active parameters; matches its previous closed Alice AI LLM with 3x fewer parameters
Released with two Russian benchmarks, WikiWebFacts and HardMultiQA; training-pipeline notes include a cascade classifier cutting data-prep compute >10x

What happened

Yandex published the base weights of a new from-scratch MoE model, not a fine-tune of Qwen or Llama. It is a small-active-parameter design meant to be cheap to run, and a preview of the architecture of Yandex's next reasoning model.

Why it matters

It is the most capable open-weights model from a Russian lab so far and adds a strong Russian-language base model under a permissive license. The benchmark claims are Yandex's own.

Changelog

2026-09-30: created (resolves the leads.md line on AliceAI-Foundation-80B-A3B-Base)

Sources (3)

id: 2026-09-21-yandex-aliceai-foundation-80b-open-weights · updated 2026-09-30 · open in the interactive timeline