DeepSeek V4 Flash 300% faster on budget GPU+CPU setup

A user achieved a 300% speedup running a 98GB quantized DeepSeek V4 Flash model (UD-Q2_K_XL) on a single RTX 4060 Ti (16GB VRAM) with a 6-core CPU, improving from 2 to 7 tokens per second. The performance gain occurred between llama.cpp versions b9986 and b10034, demonstrating significant optimization potential for running large models on budget hardware.