| <script id="data" type="application/json">{"moe": [{"evidence_id": "M21", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 1, "cache_mode": "cold", "per_stream_tok_s_median": 172.7, "aggregate_output_tok_s": 27.1, "ttft_p95_s": 7.952, "itl_p95_ms": 15.8, "kv_usage_max": 0.0438, "preemptions": 0.0}, {"evidence_id": "M22", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 4, "cache_mode": "cold", "per_stream_tok_s_median": 22.7, "aggregate_output_tok_s": 32.4, "ttft_p95_s": 29.795, "itl_p95_ms": 503.3, "kv_usage_max": 0.1752, "preemptions": 0.0}, {"evidence_id": "M23", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 8, "cache_mode": "cold", "per_stream_tok_s_median": 9.5, "aggregate_output_tok_s": 33.3, "ttft_p95_s": 59.185, "itl_p95_ms": 566.3, "kv_usage_max": 0.3504, "preemptions": 0.0}, {"evidence_id": "M24", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 12, "cache_mode": "cold", "per_stream_tok_s_median": 5.9, "aggregate_output_tok_s": 33.1, "ttft_p95_s": 88.865, "itl_p95_ms": 582.1, "kv_usage_max": 0.5256, "preemptions": 0.0}, {"evidence_id": "M25", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 16, "cache_mode": "cold", "per_stream_tok_s_median": 4.4, "aggregate_output_tok_s": 33.2, "ttft_p95_s": 120.645, "itl_p95_ms": 582.6, "kv_usage_max": 0.5256, "preemptions": 0.0}, {"evidence_id": "M26", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 24, "cache_mode": "cold", "per_stream_tok_s_median": 3.3, "aggregate_output_tok_s": 33.5, "ttft_p95_s": 172.739, "itl_p95_ms": 587.4, "kv_usage_max": 0.5256, "preemptions": 0.0}, {"evidence_id": "M27", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 32, "cache_mode": "cold", "per_stream_tok_s_median": 3.1, "aggregate_output_tok_s": 33.6, "ttft_p95_s": 234.346, "itl_p95_ms": 592.8, "kv_usage_max": 0.5256, "preemptions": 0.0}, {"evidence_id": "M28", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 1, "cache_mode": "cached", "per_stream_tok_s_median": 208.3, "aggregate_output_tok_s": 143.1, "ttft_p95_s": 0.54, "itl_p95_ms": 5.9, "kv_usage_max": 0.0197, "preemptions": 0.0}, {"evidence_id": "M29", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 4, "cache_mode": "cached", "per_stream_tok_s_median": 114.0, "aggregate_output_tok_s": 334.3, "ttft_p95_s": 1.061, "itl_p95_ms": 17.0, "kv_usage_max": 0.0271, "preemptions": 0.0}, {"evidence_id": "M30", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 8, "cache_mode": "cached", "per_stream_tok_s_median": 90.8, "aggregate_output_tok_s": 513.8, "ttft_p95_s": 1.774, "itl_p95_ms": 17.4, "kv_usage_max": 0.037, "preemptions": 0.0}, {"evidence_id": "M31", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 12, "cache_mode": "cached", "per_stream_tok_s_median": 52.5, "aggregate_output_tok_s": 467.7, "ttft_p95_s": 2.558, "itl_p95_ms": 34.0, "kv_usage_max": 0.0469, "preemptions": 0.0}, {"evidence_id": "M32", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 16, "cache_mode": "cached", "per_stream_tok_s_median": 47.0, "aggregate_output_tok_s": 442.5, "ttft_p95_s": 7.362, "itl_p95_ms": 33.8, "kv_usage_max": 0.0487, "preemptions": 0.0}, {"evidence_id": "M33", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 24, "cache_mode": "cached", "per_stream_tok_s_median": 37.0, "aggregate_output_tok_s": 413.4, "ttft_p95_s": 9.612, "itl_p95_ms": 42.1, "kv_usage_max": 0.0487, "preemptions": 0.0}, {"evidence_id": "M34", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 60000, "concurrency": 32, "cache_mode": "cached", "per_stream_tok_s_median": 37.4, "aggregate_output_tok_s": 442.2, "ttft_p95_s": 16.062, "itl_p95_ms": 35.9, "kv_usage_max": 0.0487, "preemptions": 0.0}, {"evidence_id": "M35", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 1, "cache_mode": "cold", "per_stream_tok_s_median": 172.7, "aggregate_output_tok_s": 13.4, "ttft_p95_s": 17.607, "itl_p95_ms": 14.6, "kv_usage_max": 0.0913, "preemptions": 0.0}, {"evidence_id": "M36", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 4, "cache_mode": "cold", "per_stream_tok_s_median": 10.6, "aggregate_output_tok_s": 14.4, "ttft_p95_s": 68.938, "itl_p95_ms": 668.6, "kv_usage_max": 0.3541, "preemptions": 0.0}, {"evidence_id": "M37", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 8, "cache_mode": "cold", "per_stream_tok_s_median": 4.2, "aggregate_output_tok_s": 14.6, "ttft_p95_s": 137.681, "itl_p95_ms": 720.5, "kv_usage_max": 0.7082, "preemptions": 0.0}, {"evidence_id": "M38", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 12, "cache_mode": "cold", "per_stream_tok_s_median": 2.7, "aggregate_output_tok_s": 14.7, "ttft_p95_s": 206.81, "itl_p95_ms": 735.6, "kv_usage_max": 0.8816, "preemptions": 0.0}, {"evidence_id": "M39", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 16, "cache_mode": "cold", "per_stream_tok_s_median": 2.0, "aggregate_output_tok_s": 14.7, "ttft_p95_s": 275.692, "itl_p95_ms": 740.9, "kv_usage_max": 0.8816, "preemptions": 0.0}, {"evidence_id": "M40", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 24, "cache_mode": "cold", "per_stream_tok_s_median": 1.7, "aggregate_output_tok_s": 14.8, "ttft_p95_s": 396.862, "itl_p95_ms": 744.7, "kv_usage_max": 0.8846, "preemptions": 0.0}, {"evidence_id": "M41", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 32, "cache_mode": "cold", "per_stream_tok_s_median": 1.7, "aggregate_output_tok_s": 14.8, "ttft_p95_s": 535.455, "itl_p95_ms": 745.0, "kv_usage_max": 0.8846, "preemptions": 0.0}, {"evidence_id": "M42", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 1, "cache_mode": "cached", "per_stream_tok_s_median": 175.6, "aggregate_output_tok_s": 122.0, "ttft_p95_s": 0.626, "itl_p95_ms": 14.3, "kv_usage_max": 0.0376, "preemptions": 0.0}, {"evidence_id": "M43", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 4, "cache_mode": "cached", "per_stream_tok_s_median": 119.3, "aggregate_output_tok_s": 279.5, "ttft_p95_s": 1.768, "itl_p95_ms": 17.0, "kv_usage_max": 0.045, "preemptions": 0.0}, {"evidence_id": "M44", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 8, "cache_mode": "cached", "per_stream_tok_s_median": 105.8, "aggregate_output_tok_s": 382.7, "ttft_p95_s": 3.487, "itl_p95_ms": 16.3, "kv_usage_max": 0.0524, "preemptions": 0.0}, {"evidence_id": "M45", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 12, "cache_mode": "cached", "per_stream_tok_s_median": 80.2, "aggregate_output_tok_s": 390.8, "ttft_p95_s": 5.551, "itl_p95_ms": 32.5, "kv_usage_max": 0.0568, "preemptions": 0.0}, {"evidence_id": "M46", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 16, "cache_mode": "cached", "per_stream_tok_s_median": 49.0, "aggregate_output_tok_s": 391.5, "ttft_p95_s": 7.5, "itl_p95_ms": 45.5, "kv_usage_max": 0.0648, "preemptions": 0.0}, {"evidence_id": "M47", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 24, "cache_mode": "cached", "per_stream_tok_s_median": 47.4, "aggregate_output_tok_s": 421.4, "ttft_p95_s": 10.775, "itl_p95_ms": 46.9, "kv_usage_max": 0.0666, "preemptions": 0.0}, {"evidence_id": "M48", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 120000, "concurrency": 32, "cache_mode": "cached", "per_stream_tok_s_median": 40.6, "aggregate_output_tok_s": 427.3, "ttft_p95_s": 15.945, "itl_p95_ms": 49.2, "kv_usage_max": 0.0666, "preemptions": 0.0}, {"evidence_id": "M49", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 1, "cache_mode": "cold", "per_stream_tok_s_median": 172.2, "aggregate_output_tok_s": 7.1, "ttft_p95_s": 34.578, "itl_p95_ms": 11.9, "kv_usage_max": 0.1474, "preemptions": 0.0}, {"evidence_id": "M50", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 4, "cache_mode": "cold", "per_stream_tok_s_median": 5.5, "aggregate_output_tok_s": 7.4, "ttft_p95_s": 135.521, "itl_p95_ms": 885.0, "kv_usage_max": 0.5904, "preemptions": 0.0}, {"evidence_id": "M51", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 8, "cache_mode": "cold", "per_stream_tok_s_median": 2.2, "aggregate_output_tok_s": 7.5, "ttft_p95_s": 270.78, "itl_p95_ms": 938.4, "kv_usage_max": 0.8828, "preemptions": 0.0}, {"evidence_id": "M52", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 12, "cache_mode": "cold", "per_stream_tok_s_median": 1.5, "aggregate_output_tok_s": 7.5, "ttft_p95_s": 408.007, "itl_p95_ms": 956.9, "kv_usage_max": 0.8828, "preemptions": 0.0}, {"evidence_id": "M53", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 16, "cache_mode": "cold", "per_stream_tok_s_median": 1.5, "aggregate_output_tok_s": 7.1, "ttft_p95_s": 571.368, "itl_p95_ms": 1012.3, "kv_usage_max": 0.8822, "preemptions": 0.0}, {"evidence_id": "M54", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 24, "cache_mode": "cold", "per_stream_tok_s_median": 1.3, "aggregate_output_tok_s": 6.5, "ttft_p95_s": 910.037, "itl_p95_ms": 1079.2, "kv_usage_max": 0.8846, "preemptions": 0.0}, {"evidence_id": "M55", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 32, "cache_mode": "cold", "per_stream_tok_s_median": 1.5, "aggregate_output_tok_s": 7.1, "ttft_p95_s": 1119.028, "itl_p95_ms": 1029.1, "kv_usage_max": 0.8846, "preemptions": 0.0}, {"evidence_id": "M56", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 1, "cache_mode": "cached", "per_stream_tok_s_median": 168.1, "aggregate_output_tok_s": 97.7, "ttft_p95_s": 1.084, "itl_p95_ms": 17.6, "kv_usage_max": 0.0611, "preemptions": 0.0}, {"evidence_id": "M57", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 4, "cache_mode": "cached", "per_stream_tok_s_median": 106.4, "aggregate_output_tok_s": 209.8, "ttft_p95_s": 2.835, "itl_p95_ms": 27.0, "kv_usage_max": 0.0685, "preemptions": 0.0}, {"evidence_id": "M58", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 8, "cache_mode": "cached", "per_stream_tok_s_median": 82.4, "aggregate_output_tok_s": 270.8, "ttft_p95_s": 5.362, "itl_p95_ms": 28.5, "kv_usage_max": 0.0753, "preemptions": 0.0}, {"evidence_id": "M59", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 12, "cache_mode": "cached", "per_stream_tok_s_median": 72.2, "aggregate_output_tok_s": 305.6, "ttft_p95_s": 7.795, "itl_p95_ms": 32.7, "kv_usage_max": 0.0777, "preemptions": 0.0}, {"evidence_id": "M60", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 16, "cache_mode": "cached", "per_stream_tok_s_median": 70.2, "aggregate_output_tok_s": 330.0, "ttft_p95_s": 10.137, "itl_p95_ms": 30.8, "kv_usage_max": 0.0777, "preemptions": 0.0}, {"evidence_id": "M61", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 24, "cache_mode": "cached", "per_stream_tok_s_median": 67.3, "aggregate_output_tok_s": 351.3, "ttft_p95_s": 14.707, "itl_p95_ms": 38.0, "kv_usage_max": 0.0777, "preemptions": 0.0}, {"evidence_id": "M62", "host": "rental", "phase_tag": "B", "model": "nex-n2.5-mini", "context_target_tokens": 200000, "concurrency": 32, "cache_mode": "cached", "per_stream_tok_s_median": 67.1, "aggregate_output_tok_s": 365.7, "ttft_p95_s": 19.538, "itl_p95_ms": 40.2, "kv_usage_max": 0.0777, "preemptions": 0.0}, {"evidence_id": "M64", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 1, "cache_mode": "cold", "per_stream_tok_s_median": 80.9, "aggregate_output_tok_s": 8.6, "ttft_p95_s": 26.441, "itl_p95_ms": 27.8, "kv_usage_max": 0.1047, "preemptions": 0.0}, {"evidence_id": "M65", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 4, "cache_mode": "cold", "per_stream_tok_s_median": 7.1, "aggregate_output_tok_s": 9.6, "ttft_p95_s": 103.227, "itl_p95_ms": 1990.1, "kv_usage_max": 0.4433, "preemptions": 0.0}, {"evidence_id": "M66", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 8, "cache_mode": "cold", "per_stream_tok_s_median": 2.8, "aggregate_output_tok_s": 9.7, "ttft_p95_s": 206.51, "itl_p95_ms": 2106.4, "kv_usage_max": 0.8939, "preemptions": 0.0}, {"evidence_id": "M67", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 12, "cache_mode": "cold", "per_stream_tok_s_median": 1.8, "aggregate_output_tok_s": 9.7, "ttft_p95_s": 313.661, "itl_p95_ms": 2132.9, "kv_usage_max": 0.8968, "preemptions": 0.0}, {"evidence_id": "M68", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 16, "cache_mode": "cold", "per_stream_tok_s_median": 1.4, "aggregate_output_tok_s": 9.7, "ttft_p95_s": 417.218, "itl_p95_ms": 2152.6, "kv_usage_max": 0.9099, "preemptions": 0.0}, {"evidence_id": "M69", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 24, "cache_mode": "cold", "per_stream_tok_s_median": 1.4, "aggregate_output_tok_s": 9.7, "ttft_p95_s": 604.522, "itl_p95_ms": 2164.2, "kv_usage_max": 0.9099, "preemptions": 0.0}, {"evidence_id": "M70", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 32, "cache_mode": "cold", "per_stream_tok_s_median": 1.4, "aggregate_output_tok_s": 9.7, "ttft_p95_s": 815.807, "itl_p95_ms": 2178.3, "kv_usage_max": 0.9099, "preemptions": 0.0}, {"evidence_id": "M71", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 1, "cache_mode": "cached", "per_stream_tok_s_median": 91.0, "aggregate_output_tok_s": 79.1, "ttft_p95_s": 0.41, "itl_p95_ms": 22.4, "kv_usage_max": 0.061, "preemptions": 0.0}, {"evidence_id": "M72", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 4, "cache_mode": "cached", "per_stream_tok_s_median": 65.0, "aggregate_output_tok_s": 218.4, "ttft_p95_s": 0.994, "itl_p95_ms": 34.1, "kv_usage_max": 0.0785, "preemptions": 0.0}, {"evidence_id": "M73", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 8, "cache_mode": "cached", "per_stream_tok_s_median": 48.3, "aggregate_output_tok_s": 314.2, "ttft_p95_s": 1.783, "itl_p95_ms": 37.8, "kv_usage_max": 0.1017, "preemptions": 0.0}, {"evidence_id": "M74", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 12, "cache_mode": "cached", "per_stream_tok_s_median": 34.1, "aggregate_output_tok_s": 317.0, "ttft_p95_s": 2.986, "itl_p95_ms": 53.5, "kv_usage_max": 0.1294, "preemptions": 0.0}, {"evidence_id": "M75", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 16, "cache_mode": "cached", "per_stream_tok_s_median": 35.9, "aggregate_output_tok_s": 321.2, "ttft_p95_s": 9.027, "itl_p95_ms": 49.0, "kv_usage_max": 0.1294, "preemptions": 0.0}, {"evidence_id": "M76", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 24, "cache_mode": "cached", "per_stream_tok_s_median": 30.3, "aggregate_output_tok_s": 322.3, "ttft_p95_s": 12.144, "itl_p95_ms": 53.6, "kv_usage_max": 0.1294, "preemptions": 0.0}, {"evidence_id": "M77", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 60000, "concurrency": 32, "cache_mode": "cached", "per_stream_tok_s_median": 29.6, "aggregate_output_tok_s": 318.7, "ttft_p95_s": 20.561, "itl_p95_ms": 53.9, "kv_usage_max": 0.125, "preemptions": 0.0}, {"evidence_id": "M78", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 1, "cache_mode": "cold", "per_stream_tok_s_median": 87.4, "aggregate_output_tok_s": 3.9, "ttft_p95_s": 62.549, "itl_p95_ms": 15.3, "kv_usage_max": 0.1948, "preemptions": 0.0}, {"evidence_id": "M79", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 4, "cache_mode": "cold", "per_stream_tok_s_median": 3.3, "aggregate_output_tok_s": 4.5, "ttft_p95_s": 225.014, "itl_p95_ms": 2358.7, "kv_usage_max": 0.8706, "preemptions": 0.0}, {"evidence_id": "M80", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 8, "cache_mode": "cold", "per_stream_tok_s_median": 1.5, "aggregate_output_tok_s": 4.5, "ttft_p95_s": 453.286, "itl_p95_ms": 2453.7, "kv_usage_max": 0.8968, "preemptions": 0.0}, {"evidence_id": "M81", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 12, "cache_mode": "cold", "per_stream_tok_s_median": 1.5, "aggregate_output_tok_s": 4.5, "ttft_p95_s": 683.317, "itl_p95_ms": 2466.4, "kv_usage_max": 0.8924, "preemptions": 0.0}, {"evidence_id": "M82", "host": "rental", "phase_tag": "QB", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 16, "cache_mode": "cold", "per_stream_tok_s_median": 1.5, "aggregate_output_tok_s": 4.5, "ttft_p95_s": 913.008, "itl_p95_ms": 2486.5, "kv_usage_max": 0.8968, "preemptions": 0.0}, {"evidence_id": "M90", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 32, "cache_mode": "cold", "per_stream_tok_s_median": 3.0, "aggregate_output_tok_s": 8.9, "ttft_p95_s": 891.275, "itl_p95_ms": 1374.8, "kv_usage_max": 0.8955, "preemptions": 0.0}, {"evidence_id": "M91", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 1, "cache_mode": "cached", "per_stream_tok_s_median": 80.4, "aggregate_output_tok_s": 68.7, "ttft_p95_s": 0.531, "itl_p95_ms": 25.2, "kv_usage_max": 0.1161, "preemptions": 0.0}, {"evidence_id": "M92", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 4, "cache_mode": "cached", "per_stream_tok_s_median": 61.8, "aggregate_output_tok_s": 199.2, "ttft_p95_s": 1.349, "itl_p95_ms": 30.0, "kv_usage_max": 0.1335, "preemptions": 0.0}, {"evidence_id": "M93", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 8, "cache_mode": "cached", "per_stream_tok_s_median": 47.9, "aggregate_output_tok_s": 296.6, "ttft_p95_s": 2.402, "itl_p95_ms": 35.6, "kv_usage_max": 0.1567, "preemptions": 0.0}, {"evidence_id": "M94", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 12, "cache_mode": "cached", "per_stream_tok_s_median": 39.0, "aggregate_output_tok_s": 243.5, "ttft_p95_s": 8.98, "itl_p95_ms": 41.2, "kv_usage_max": 0.2351, "preemptions": 0.0}, {"evidence_id": "M95", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 16, "cache_mode": "cached", "per_stream_tok_s_median": 36.2, "aggregate_output_tok_s": 303.5, "ttft_p95_s": 9.673, "itl_p95_ms": 42.9, "kv_usage_max": 0.1843, "preemptions": 0.0}, {"evidence_id": "M96", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 24, "cache_mode": "cached", "per_stream_tok_s_median": 32.5, "aggregate_output_tok_s": 357.3, "ttft_p95_s": 11.539, "itl_p95_ms": 42.8, "kv_usage_max": 0.1843, "preemptions": 0.0}, {"evidence_id": "M97", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 120000, "concurrency": 32, "cache_mode": "cached", "per_stream_tok_s_median": 32.6, "aggregate_output_tok_s": 343.0, "ttft_p95_s": 18.877, "itl_p95_ms": 42.9, "kv_usage_max": 0.1843, "preemptions": 0.0}, {"evidence_id": "M98", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 200000, "concurrency": 1, "cache_mode": "cold", "per_stream_tok_s_median": 75.7, "aggregate_output_tok_s": 4.4, "ttft_p95_s": 55.158, "itl_p95_ms": 13.9, "kv_usage_max": 0.3251, "preemptions": 0.0}, {"evidence_id": "M99", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 200000, "concurrency": 4, "cache_mode": "cold", "per_stream_tok_s_median": 7.3, "aggregate_output_tok_s": 4.4, "ttft_p95_s": 231.44, "itl_p95_ms": 1737.4, "kv_usage_max": 0.9942, "preemptions": 2.0}, {"evidence_id": "M100", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 200000, "concurrency": 8, "cache_mode": "cold", "per_stream_tok_s_median": 3.7, "aggregate_output_tok_s": 4.3, "ttft_p95_s": 468.193, "itl_p95_ms": 1800.6, "kv_usage_max": 0.9942, "preemptions": 2.0}, {"evidence_id": "M101", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 200000, "concurrency": 1, "cache_mode": "cached", "per_stream_tok_s_median": 75.9, "aggregate_output_tok_s": 61.1, "ttft_p95_s": 0.805, "itl_p95_ms": 26.8, "kv_usage_max": 0.1901, "preemptions": 0.0}, {"evidence_id": "M102", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 200000, "concurrency": 4, "cache_mode": "cached", "per_stream_tok_s_median": 55.3, "aggregate_output_tok_s": 164.8, "ttft_p95_s": 2.085, "itl_p95_ms": 33.0, "kv_usage_max": 0.2119, "preemptions": 0.0}, {"evidence_id": "M103", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 200000, "concurrency": 8, "cache_mode": "cached", "per_stream_tok_s_median": 38.7, "aggregate_output_tok_s": 227.5, "ttft_p95_s": 3.618, "itl_p95_ms": 43.6, "kv_usage_max": 0.2308, "preemptions": 0.0}, {"evidence_id": "M104", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 200000, "concurrency": 12, "cache_mode": "cached", "per_stream_tok_s_median": 30.9, "aggregate_output_tok_s": 265.0, "ttft_p95_s": 5.258, "itl_p95_ms": 51.7, "kv_usage_max": 0.2583, "preemptions": 0.0}, {"evidence_id": "M105", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 200000, "concurrency": 16, "cache_mode": "cached", "per_stream_tok_s_median": 28.3, "aggregate_output_tok_s": 243.0, "ttft_p95_s": 12.533, "itl_p95_ms": 52.2, "kv_usage_max": 0.2583, "preemptions": 0.0}, {"evidence_id": "M106", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 200000, "concurrency": 24, "cache_mode": "cached", "per_stream_tok_s_median": 25.8, "aggregate_output_tok_s": 275.9, "ttft_p95_s": 15.541, "itl_p95_ms": 52.7, "kv_usage_max": 0.2583, "preemptions": 0.0}, {"evidence_id": "M107", "host": "rig", "phase_tag": "QB-rig", "model": "qwen3.8-27b", "context_target_tokens": 200000, "concurrency": 32, "cache_mode": "cached", "per_stream_tok_s_median": 25.3, "aggregate_output_tok_s": 269.6, "ttft_p95_s": 24.496, "itl_p95_ms": 54.0, "kv_usage_max": 0.2583, "preemptions": 0.0}], "flash": [{"workload": "full window: ~254K prompt, 8,192 output tokens, one request", "profile": "Frozen EXL", "quantization": "EXL3 2.50", "cache": "Quant8", "gpus": 2, "mode": "cold", "prompt_tokens": 253952, "output_tokens": 8192, "wall_tok_s": 131.50802990523357, "wall_seconds": 62.29277410591021, "ttft_seconds": 38.756553784012794}, {"workload": "full window: ~254K prompt, 8,192 output tokens, one request", "profile": "Frozen EXL", "quantization": "EXL3 2.50", "cache": "Quant8", "gpus": 2, "mode": "warm", "prompt_tokens": 253952, "output_tokens": 8192, "wall_tok_s": 341.3037495659041, "wall_seconds": 24.00208028894849, "ttft_seconds": 1.4291184929898009}, {"workload": "full window: ~254K prompt, 8,192 output tokens, one request", "profile": "Tuned EXL", "quantization": "EXL3 2.50", "cache": "Quant8", "gpus": 2, "mode": "cold", "prompt_tokens": 253952, "output_tokens": 8192, "wall_tok_s": 136.67608003388736, "wall_seconds": 59.93733503308613, "ttft_seconds": 38.52612456004135}, {"workload": "full window: ~254K prompt, 8,192 output tokens, one request", "profile": "Tuned EXL", "quantization": "EXL3 2.50", "cache": "Quant8", "gpus": 2, "mode": "warm", "prompt_tokens": 253952, "output_tokens": 8192, "wall_tok_s": 371.9796256329185, "wall_seconds": 22.022711555939168, "ttft_seconds": 1.4781194840325043}, {"workload": "full window: ~254K prompt, 8,192 output tokens, one request", "profile": "Strata serial", "quantization": "GSQ IQ3_XXS", "cache": "INT8", "gpus": 2, "mode": "cold", "prompt_tokens": 253943, "output_tokens": 8192, "wall_tok_s": 147.97840141069352, "wall_seconds": 55.35943030810449, "ttft_seconds": 27.562829514034092}, {"workload": "full window: ~254K prompt, 8,192 output tokens, one request", "profile": "Strata serial", "quantization": "GSQ IQ3_XXS", "cache": "INT8", "gpus": 2, "mode": "warm", "prompt_tokens": 253943, "output_tokens": 8192, "wall_tok_s": 292.1141084845735, "wall_seconds": 28.043835480930284, "ttft_seconds": 1.5177656019805}, {"workload": "full window: ~254K prompt, 8,192 output tokens, one request", "profile": "NInfer-ext", "quantization": "custom EXL3 3.5", "cache": "FP8", "gpus": 1, "mode": "cold", "prompt_tokens": 253952, "output_tokens": 8192, "wall_tok_s": 28.208987305144934, "wall_seconds": 290.40390253590886, "ttft_seconds": 101.90239312092308}, {"workload": "full window: ~254K prompt, 8,192 output tokens, one request", "profile": "Stock llama.cpp", "quantization": "GSQ IQ3_XXS", "cache": "q8_0", "gpus": 2, "mode": "cold", "prompt_tokens": 253950, "output_tokens": 8192, "wall_tok_s": 20.250885298796316, "wall_seconds": 404.5255246439483, "ttft_seconds": 163.5513619859703}, {"workload": "full window: ~254K prompt, 8,192 output tokens, one request", "profile": "BeeLlama", "quantization": "GSQ IQ3_XXS", "cache": "q8_0", "gpus": 2, "mode": "cold", "prompt_tokens": 253950, "output_tokens": 8192, "wall_tok_s": 17.812060727053534, "wall_seconds": 459.9130962740164, "ttft_seconds": 213.1621092199348}], "models": {"nex-n2.5-mini": "ProCreations/Nex-N2.5-mini-NVFP4", "qwen3.8-27b": "unsloth/Qwen3.8-27B-NVFP4"}, "links": {"moe_ds": "https://huggingface.co/datasets/Scalably/moe-vs-dense-vllm-rtx5090", "flash_ds": "https://huggingface.co/datasets/Scalably/qwen-flash-dual-5090-serving", "moe_article": "https://scalably.io/blog/moe-vs-dense-llm-inference-vllm-benchmark-rtx-5090", "flash_article": "https://scalably.io/blog/qwen-flash-next-dual-rtx-5090-serving-benchmark", "org": "https://huggingface.co/Scalably"}}</script> |
