yandex/AliceAI-Foundation-80B-A3B-Base config.json

config.json fetched 2026-10-18. model_type alice_ai with custom modeling via auto_map (configuration_alice_ai.AliceAIConfig / modeling_alice_ai.AliceAIForCausalLM), so it does not load in stock transformers without the repo code and trust_remote_code. The card pins transformers 5.16.1. The hybrid split (36 linear_attention + 12 full_attention), the MoE routing (512 experts, 10 per token, shared expert of intermediate size 512) and the gated-attention geometry (16 query heads, 2 kv heads, head_dim 256) all come from this file.