How to speed up inference on DGX Spark?

A DGX Spark user running the granite-4.2-30b-fp8 model reported getting only 4-6 output tokens per second with llama.cpp and vLLM, and reached 40 tokens per second after a forum respondent identified as lambda advised switching to a mixture-of-experts model, specifically Qwen 3.8 Flash Next. The exchange, posted on the Level1Techs forum, illustrates that model choice rather than inference engine tuning drove the roughly 7-10x throughput gain on the hardware.