@berryxia: Damn, even my eyes can't keep up with this speed! Daniel Han, founder of UnslothAI, YC S24, previously at NVIDIA doing ML, just released the experimental MTP GGUF of Qwen3.6. The 27B model hits 140 tokens/s on a single GPU. 35B-A...

Damnnn, even my eyes can't keep up with this speed! Daniel Han, founder of UnslothAI, YC S24, previously at NVIDIA doing ML, just released the experimental MTP GGUF of Qwen3.6. The 27B model runs at 140 tokens/s on a single GPU. The 35B-A3B version is even more insane, hitting 220 tokens/s. That's over 1.4x faster than the original GGUF, with zero accuracy loss. They tested extensively and found that setting draft tokens to 2 is the sweet spot — any higher and the acceptance rate plummets, causing actual speed to drop. Looking at that benchmark curve, my biggest takeaway is that the performance ceiling for local large models has been pushed up significantly. I used to think 30B+ models were too slow to run locally, but now MTP speculative decoding is squeezing every drop of potential out of consumer GPUs. If you're playing with llama.cpp, running local agents, or doing daily coding, you need to try this update immediately. Local AI is starting to feel less and less like a "compromise version."