GPT-OSS 120B boosts throughput on DGX Spark

This looks more like a stack mismatch than a hard hardware ceiling. GPT-OSS 120B is sparse, open-weight, and native MXFP4, so the fastest path is usually to respect the model's format and let the runtime/kernel stack do the heavy lifting.

• –OpenAI says GPT-OSS 120B has 117B total parameters but only 5.1B active per token, which means kernel efficiency matters a lot.
• –The thread's own numbers line up with that story: ~32 tps in vLLM/Q4_K_S, roughly ~50 tps after switching to llama.cpp/MXFP4, and one reply expecting around 60 tps on DGX Spark.
• –NVIDIA's DGX Spark blog says llama.cpp optimizations have lifted performance by about 35% on average, reinforcing that runtime choice is the biggest lever.
• –If you care about response quality, the win is native precision plus flash-attn, batching, and context tuning, not a harsher quant.