Testing MTP/Speculative Decoding with latest LLAMA.cpp shows that with a vision model like Qwen3.6-27B-MTP There is no gain in token generation with the vision mmproj loaded. If you don't enable vision by not passing the mmproj then you benefit from the MTP/Speculative decoding boost.
Ideally a way create a profile before loading a model or a way via the side menu to disable vision when using --spec-type draft-mtp
Testing MTP/Speculative Decoding with latest LLAMA.cpp shows that with a vision model like Qwen3.6-27B-MTP There is no gain in token generation with the vision mmproj loaded. If you don't enable vision by not passing the mmproj then you benefit from the MTP/Speculative decoding boost.
Ideally a way create a profile before loading a model or a way via the side menu to disable vision when using --spec-type draft-mtp