NVIDIA stacks three optimizations to speed MoE inference on DGX Spark

Key points

• Three optimizations applied together run MoE inference on a single DGX Spark machine.
• Ant Group inclusionAI's 124B Ling-3.0-flash model saw GSM8K decode throughput rise 42.5%.
• No detectable accuracy change was reported on the GSM8K benchmark.

Why it matters for marketers: Teams running large mixture-of-experts models on a single workstation-class system may get faster decoding without a measured accuracy cost.