Tag
Benchmark tests on Ling-3.0-flash show that higher acceptance length in multi-token prediction decreases prose throughput, with n=1 being the most efficient setting for the evaluated workloads.