ResearchBenchmarksMath Reasoning on Minerva (Latency and Throughput)Math Reasoning on Minerva (Latency and Throughput)Follow3.65Tau (Accepted Tokens)LongSpec0.8941.60952.3253.0405Feb 24, 2025Evaluation resultsMethodMethodLinksTau (Accepted Tokens)Tokens/sRelative Improvement (x)LongSpecModel=QwQ-32B, Maximum...Model=QwQ-32B, Maximum output length=32k tokens2025.023.6544.512.29VanillaModel=QwQ-32B, Maximum...Model=QwQ-32B, Maximum output length=32k tokens2025.02119.46—