NVIDIA lifts Ling-3.0-flash GSM8K decode throughput 42.5% on a single DGX Spark with 3 MoE optimizations

NVIDIA has shown how 3 stacked optimizations speed up Mixture-of-Experts (MoE) inference on a single DGX Spark. On Ant Group inclusionAI's 124B Ling-3.0-flash, GSM8K decode throughput rose 42.5% with no detectable accuracy change, according to a MarkTechPost report published November 3, 2026.

Key figures

• Hardware: a single NVIDIA DGX Spark
• Model: Ling-3.0-flash from Ant Group inclusionAI, 124B parameters
• Benchmark: GSM8K, decode throughput
• Result: 42.5% higher throughput
• Accuracy: no detectable change

The optimizations are described as stacked, meaning they are applied together to reach the reported gain. The 42.5% figure is the result NVIDIA reports for the combined set on this model and benchmark.

What this means

The result is a decode throughput gain on a 124B MoE model running on one DGX Spark, measured on GSM8K. The report states accuracy did not change in any detectable way, so the speedup was not traded against GSM8K quality. The 42.5% figure applies to the model, hardware and benchmark named above.