ResearchBenchmarksMath Reasoning on AIME 24 (Token Usage and Throughput)Math Reasoning on AIME 24 (Token Usage and Throughput)Follow3.82Mean Accepted TokensLongSpec0.88721.64862.413.1714Feb 24, 2025Evaluation resultsMethodMethodLinksMean Accepted TokensThroughput (Tokens/s)Improvement FactorLongSpecModel=QwQ-32B, Maximum...Model=QwQ-32B, Maximum output length=32k tokens2025.023.8242.632.25VanillaModel=QwQ-32B, Maximum...Model=QwQ-32B, Maximum output length=32k tokens2025.02118.92—