Community
Betelgeuse runs Qwen 3.8 Flash Next at 22 tok/s on an RTX 3060 with 16GB RAM
A community fork of llama.cpp reports 22 tok/s sustained, with bursts to 48 tok/s, using a 68GB IQ2_XS build without gate pruning. The author reports bit-exact output versus stock llama.cpp in their tests. Results are self-reported and not independently verified.