ResearchBenchmarksDecode Throughput on Gemma 26B-A4B 4Follow69.3Throughput (tok/s)MLX57.54860.59963.6566.701Jul 1, 2026Evaluation ResultsMethodMethodLinksThroughput (tok/s)Speedup vs. llama.cppRatio vs. MLXMLXQuantization=Q4, Hardw...Quantization=Q4, Hardware=Apple M4 Pro, Target Length=tg1282026.0769.3——BaseRTQuantization=Q4, Hardw...Quantization=Q4, Hardware=Apple M4 Pro, Target Length=tg1282026.0762.21.070.9llama.cppQuantization=Q4, Hardw...Quantization=Q4, Hardware=Apple M4 Pro, Target Length=tg1282026.0758——