Qwen3 30B (3B active)

User reports 50-54 t/s with a custom CUDA/C++ engine, a 50% improvement over llama.cpp at 33-34 t/s.