ResearchBenchmarksLatency and Throughput on MultiNews, Qasper, RepoBench-P, and RULER Averaged 128K (test)Follow15.3Memory Footprint (GB)TTKV12.23232.94153.6574.359Mar 27, 2026Evaluation ResultsMethodMethodLinksMemory Footprint (GB)Average Latency (ms)TTKVModel=Llama-3.1-70BModel=Llama-3.1-70B2026.0315.3595TTKVModel=Qwen2.5-32BModel=Qwen2.5-32B2026.0315.5555TTKVModel=DeepSeek-R1-14BModel=DeepSeek-R1-14B2026.0315.8465FP16Model=Llama-3.1-70BModel=Llama-3.1-70B2026.03922,450FP16Model=Qwen2.5-32BModel=Qwen2.5-32B2026.03922,350FP16Model=DeepSeek-R1-14BModel=DeepSeek-R1-14B2026.03922,550