[Benchmark] Nanbeige33B on RTX 5090: DFlash2 at 8K and 180K context
Nanbeige4.3-33B-EXL3-3.55bpwBoth current presets keep DFlash2 enabled: 116.9 tok/s at 8K capacity; 115.1 tok/s on the short workload with a 180224-token Q4 cache; 23.4 tok/s at 179961 occupied tokens. Raw checks and failed larger trials retained; historical480K/no-draft results are labeled separately.
5 speed tests