Model Details
AliceAI-Foundation-80B-A3B-Base
-8957e5){kind=icon}
Quick answer: AliceAI-Foundation-80B-A3B-Base is Yandex's open pre-trained (base) language model: 80B total and 3B active parameters, a hybrid KDA/gated-attention MoE with 512 experts, a 262,144-token context, trained entirely from scratch and released September 12, 2026 under Apache 2.0. Yandex reports particularly strong Russian factual knowledge (WikiWebFacts 86.5, HardMultiQA 67.9) and 91.1 on MATH-500.
At a Glance
Where it leads: Russian-language factual knowledge and exams (WikiWebFacts 86.5, CultCat 86.5, EGE CoT 90.5) and math (MATH-500 91.1) among the open base models Yandex compared.
Where it lags: English trivia (TriviaQA 79.0 vs 89.8 for Nemotron-3-Super-120B-A12B-Base) and SuperGPQA (44.3 vs 46.6).
Best for: Research and fine-tuning for Russian-language assistants; it is a base model with no post-training or alignment.
What AliceAI-Foundation-80B-A3B-Base Is
Yandex built a new training corpus, chose the architecture and hyperparameters through a series of 2-trillion-token from-scratch runs, and prepared data for complex reasoning and tool use. The 48-layer network repeats a pattern of three KDA-MoE layers followed by one gated-attention-MoE layer, with 512 experts (top-10 routed plus one shared) and a 1-layer MTP head.
The model is a pre-trained checkpoint intended for research, experimentation and further tuning rather than direct use in products. Alongside the weights Yandex published two Russian factual benchmarks, WikiWebFacts and HardMultiQA, with their evaluation protocols.
Specifications
| Field | Value |
|---|---|
| Organization | Yandex |
| Hugging Face | yandex/AliceAI-Foundation-80B-A3B-Base |
| Architecture | Hybrid KDA + gated-attention MoE, 80B total / 3B active |
| Languages | Russian, English |
| Context length | 262,144 tokens |
| License | Apache 2.0 |
| Release date | 2026-09-12 |
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| WikiWebFacts | 86.5 | Model card | 2026-09 |
| HardMultiQA | 67.9 | Model card | 2026-09 |
| MMLU-Pro (5-shot CoT, base) | 66.8 | Model card | 2026-09 |
| SuperGPQA (5-shot CoT, base) | 44.3 | Model card | 2026-09 |
| MATH-500 (5-shot, base) | 91.1 | Model card | 2026-09 |
Self-reported by Yandex on a pre-trained base model; not independent Benchgen measurements. Also reported but not added: Yandex-internal benchmarks (CultCat, EduBench, ExpertFactsQA, EGE CoT, FinQA 128k, LongMemEval 128k), TriviaQA (LLM-as-judge instead of exact match), BigCodeBench (Yandex's own implementation with improved tests), LiveCodeBench v5-6 (not v6) and pass@k reasoning results (AIME 2026 pass@32 96.7, HMMT 2026 Feb pass@32 96.9, IMO AnswerBench pass@8 88.7).