If your GGUF has MTP/NextN tensors baked in (GLM-5.2, hy_v3, qwen35moe, step35, etc.), recent llama.cpp builds load them by default — even if you never pass --spec-type draft-mtp . Before, they were skipped unless you actually enabled speculative decoding. Most community GGUFs bundle the MTP block by default, so this means extra VRAM/RAM use (~1 extra MoE layer) on every load, whether you use MTP or not. See https://github.com/ggml-org/llama.cpp/pull/25980 submitted by /...