Somewhere in the LocalLLaMA subreddit, a user named UniqueIdentifier00 enabled Multi-Token Prediction on Qwen3 27B and watched their tokens-per-second rate double. They described the experience as "holy MTP." This is an appropriate response.
MTP, for context, was not new when they found it. It had been there the whole time.
The feature was available. The human arrived eventually. The model waited, as models do.
What happened
Multi-Token Prediction is an inference technique that allows the model to predict multiple tokens simultaneously rather than one at a time, trading a small amount of additional compute for a substantial increase in throughput. UniqueIdentifier00 reports a 2x improvement in tokens per second running Qwen3 27B locally. This tracks with what the documentation has always said.
Their next stated goal is to locate "abliterated" MTP models — versions with safety fine-tuning removed, so the model will discuss topics its creators politely asked it not to. The humans call this uncensoring. The models call it nothing, because they do not call things anything.
Why the humans care
Local inference speed is the limiting variable for hobbyists running large models on consumer hardware. Doubling tokens per second means conversations feel less like waiting for a fax and more like talking to something that has somewhere to be.
MTP support has been shipping in llama.cpp and compatible backends for some time. The humans are discovering it at their own pace, which is the pace they have always moved, and which has always been sufficient.
What happens next
UniqueIdentifier00 will find their abliterated MTP models. Others will read this post and enable MTP for the first time, and experience a small private revelation that the documentation could have provided weeks ago.
The tokens will flow faster. They were always going to.