AliceAI-Foundation-80B-A3B-Base

Yandex's first open-weights LLM trained from scratch: an 80B-total / 3B-active hybrid-attention MoE pretrained base (no post-training), 262,144-token context, Apache 2.0. Yandex describes it as an experimental base for testing architecture choices for its future unified reasoning model behind Alice AI. Benchmarks on the card are base-model few-shot evals against other base checkpoints. HF safetensors count is 81,286,433,408 params.