New AI Speed Hack Makes Chatbots Answer Twice as Fast

Inco AI just shipped DFlash 2, a tool that makes AI chatbots and coding assistants respond much faster. It works with llama.cpp, a widely used program for running open-source AI models on your own computer. In a careful three-day test with the Qwen 3.8 27B model, DFlash 2 doubled the speed on 100 real coding problems — from about 68 words per second to 154. That means the time between each word dropped from 14 milliseconds to 6, so answers feel snappier and less like a paused video.

The secret is a trick called speculative decoding. Instead of the big AI predicting one word at a time, a small "draft" model suggests several possible next words. The big model then checks them all at once, which is faster than writing each word separately. DFlash 2 is that drafter, and it was tuned specifically for Qwen 3.8 27B. In one long 18-turn coding session, adding an extra lookup table pushed speeds to 4.68 times faster. But that trick only helps with coding — on regular prose it actually made things 30% slower.

Before you get too excited, the hardware matters. The test used a $6,000+ RTX PRO 6000 GPU with 96GB of memory, and DFlash 2 itself adds about 2.7GB of VRAM. So this is aimed at developers running serious AI workloads, not casual chatbot users. The team also flagged that some flashy 8x numbers were misleading benchmark artifacts, not real-world gains. Still, for anyone running coding assistants locally, doubling speed without changing the model is a big deal — it means less waiting, more productivity, and lower costs per answer.