Recent llama.cpp builds now load MTP/NextN tensors by default for draft-mtp architectures (e.g., GLM-5.2, hy_v3, qwen35moe, step35) even when speculative decoding is disabled via --spec-type. Previously these tensors were skipped unless MTP was explicitly enabled. Since most community GGUFs bundle the MTP block, this change increases VRAM/RAM usage by approximately one extra MoE layer on every model load. See PR #25980 for details.