Hunyuan-A13B model card and technical report link (huggingface.co/tencent)

Hunyuan-A13B card states 13B active out of 80B total. config.json gives 64 experts with moe_topk 8 per layer, num_shared_expert 1, 32 layers, hidden 4096, 32 heads / 8 kv heads, head_dim 128, tie_word_embeddings true, dynamic rope scaling. Note the name 'A13B' refers to active params, not total. The card also states the model supports 256K context but config.json max_position_embeddings is 32768.