Community
Betelgeuse runs Qwen 3.8 Flash Next IQ4_XS at 24 tok/s on 16GB RAM and an RTX 3060
Self-reported results show 24 tok/s with dual-channel 2x8GB DDR4 and 14 tok/s on a single 16GB stick, with output byte-exact to stock llama.cpp. The project also says early DeepSeek V4.1 Flash tests reach about 25 tok/s on similar hardware.