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.

I have tried llama.cpp and vllm but so far only getting 4-6 output tokens/seconds with granite-4.2-30b-fp8 Are there any known tips to get faster output? different models, options, bespoke patches? 1 Like lambda https://forum.level1techs.com/u/lambda 2 Pick a different model? Ideally MoE. Check out Qwen 3.8 Flash Next - 2 Likes Thank you, I was able to get 40 tokens/seconds with this