Llama.cpp MTP Support Hits Beta

This is the kind of runtime upgrade that matters more than another small kernel tweak: if MTP lands cleanly, llama.cpp gets a real path to vLLM-class decode speed without giving up its local-first appeal.

• –The reported numbers are strong enough to matter operationally, not just academically: ~83.8s wall time versus 201.07s baseline in the author’s benchmark set.
• –Packaging MTP inside the same GGUF lowers friction versus draft-model speculative decoding, which should make adoption simpler for local and self-hosted deployments.
• –The current scope is still narrow, centered on Qwen3.6 27B and MTP-enabled models, so the real test is how quickly other model families and backends follow.
• –Together with llama.cpp’s tensor-parallel progress, this could erase a big chunk of the token-generation lead that larger inference stacks have held.
• –The PR is still draft-stage, so the main question now is stability across workloads, not whether the raw speedup is real.