ResearchBenchmarksLong-Context Generation Efficiency on Qwen2.5-72BFollow4.76Speedup (8K)Dustin3.8764.10554.3354.5645Jun 23, 2026Evaluation ResultsMethodMethodLinksSpeedup (8K)Speedup (16K)Speedup (32K)DustinBatch Size=16Batch Size=162026.064.767.069.17SpecAttnBatch Size=16Batch Size=162026.063.915.627.11