ResearchBenchmarksDecode Throughput on Qwen3 30B-A3BFollow84.1Throughput (tokens/s)BaseRT80.56481.48282.483.318Jul 1, 2026Evaluation ResultsMethodMethodLinksThroughput (tokens/s)Speedup Ratio vs. llama.cppSpeedup Ratio vs. MLXBaseRTQuantization=Q4, Hardw...Quantization=Q4, Hardware=Apple M4 Pro, Target Length=tg1282026.0784.11.041.01MLXQuantization=Q4, Hardw...Quantization=Q4, Hardware=Apple M4 Pro, Target Length=tg1282026.0783.1——llama.cppQuantization=Q4, Hardw...Quantization=Q4, Hardware=Apple M4 Pro, Target Length=tg1282026.0780.7——