A speedup with no extra hardware cost and no downside is the rarest thing in local inference. Multi Token Prediction gave heise between 25 and over 250 percent more tokens per second across four models. The floor of 25 percent is the number I care about. Ceilings are marketing; floors are what you plan around.
So the move is model selection, not a shopping list. MTP isn’t available for every local model, but Gemma 4, Qwen3.5, Qwen3.6 and Qwen3.8 have it, in variants small enough for a 16 GByte card. If I’m picking something to self-host this week, MTP support goes in the column next to quality — then run llama.cpp both ways on my own prompts and measure the delta myself.
The story — heise tested Multi Token Prediction, which makes some local LLMs generate faster without new hardware. Four models produced between 25 and over 250 percent more tokens per second than without MTP. Not every model supports it; Gemma 4, Qwen3.5, Qwen3.6 and Qwen3.8 do, in sizes that run on 16 GByte cards. Testing used a Radeon RX 9060 XT and an RTX 3090 with llama.cpp, comparing two prompts and tuning a parameter (Source).