
⚡ NVIDIA accelerates Ling-3.0-flash inference by 42.5% on DGX Spark Throughput on GSM8K increased from 37.81 to 53.89 tokens per second. Combined Humming cores, an LM head in FP8, and speculative decoding. Accuracy decreased from 96.89% to 96.82%, a difference described as statistically undetectable.


