yandex/AliceAI-Foundation-80B-A3B-Base — new model trending #30 on Hugging Face

Yandex open-sources AliceAI-Foundation-80B-A3B-Base, a hybrid MoE base model with 3B active parameters and 262k context.

Yandex released AliceAI-Foundation-80B-A3B-Base, an autoregressive base language model trained from scratch with 80B total parameters and 3B active per token, a 262,144-token context, and a hybrid architecture mixing KDA (Kimi Delta Attention-style) and gated attention layers with a 512-expert MoE (top-10 plus 1 shared expert) and one MTP layer. Key architectural and data decisions were validated in separate from-scratch runs of 2 trillion tokens each. It claims results on par with larger open-source models on math and coding, and posts 86.5 on WikiWebFacts and 67.9 on HardMultiQA, beating Qwen3.5-35B-A3B-Base, GLM-4.5-Air-Base, Nemotron-3-Super-120B-A12B-Base, and DeepSeek-V4-Flash-Base, with particular strength in Russian factual knowledge. Yandex also released the WikiWebFacts and HardMultiQA Russian-language factual benchmark datasets and evaluation protocols alongside the weights.