{ "release": "Winnow-12B", "date": "2026-09-21", "local_comparison_gpu": "NVIDIA RTX PRO 5000 Blackwell", "measurements": [ { "model": "Winnow-12B Q8", "precision": "Q8_0", "quality_percent": { "jevbench": 85.71428571428571, "kev_v9": 81.54875717017208, "kev_additions": 69.23076923076923, "typed": 70.0 }, "counts": { "jevbench": 231, "kev_v9": 1046, "kev_additions": 390, "typed": 2000 }, "calibration": { "jev_brier": 0.20571291282072968, "jev_ece": 0.0741753092521077, "kev_brier": 0.2855102865874103, "kev_ece": 0.09795411445352069, "typed_teacher_brier_non_score": 0.22546611415467793, "typed_ece": 0.15708168396777994 }, "memory_gib": { "gpu": "NVIDIA RTX PRO 5000 Blackwell", "device_baseline_mib": 2.0, "load_vram_gib": 12.82421875, "idle_vram_gib": 14.333984375, "inference_vram_gib": 14.35546875, "load_ram_pss_gib": 12.184748649597168, "idle_ram_pss_gib": 0.541377067565918, "inference_ram_pss_gib": 0.5913925170898438 }, "performance": { "performance-short": { "1": { "warm": { "count": 10, "median_ms": 22.109638899564743, "p95_ms": 33.82298490032554, "range_ms": [ 21.54988795518875, 33.82298490032554 ] }, "cold": { "count": 10, "median_ms": 52.6955034583807, "p95_ms": 59.704908169806004, "range_ms": [ 51.540270913392305, 59.704908169806004 ] }, "warm_median_decisions_per_second": 45.2291421195344 }, "8": { "warm": { "count": 10, "median_ms": 92.9789764340967, "p95_ms": 94.17951526120305, "range_ms": [ 90.98931262269616, 94.17951526120305 ] }, "cold": { "count": 10, "median_ms": 121.38708005659282, "p95_ms": 127.18812981620431, "range_ms": [ 120.11612579226494, 127.18812981620431 ] }, "warm_median_decisions_per_second": 86.04095578176626 }, "16": { "warm": { "count": 10, "median_ms": 187.76516010984778, "p95_ms": 190.6993198208511, "range_ms": [ 185.49408204853535, 190.6993198208511 ] }, "cold": { "count": 10, "median_ms": 220.4458403866738, "p95_ms": 225.0557690858841, "range_ms": [ 218.5507151298225, 225.0557690858841 ] }, "warm_median_decisions_per_second": 85.21282644042995 }, "32": { "warm": { "count": 10, "median_ms": 367.638946743682, "p95_ms": 372.3510168492794, "range_ms": [ 365.28318701311946, 372.3510168492794 ] }, "cold": { "count": 10, "median_ms": 399.432020727545, "p95_ms": 411.3255850970745, "range_ms": [ 398.2086409814656, 411.3255850970745 ] }, "warm_median_decisions_per_second": 87.04192056754643 }, "64": { "warm": { "count": 10, "median_ms": 736.4177405834198, "p95_ms": 743.015272077173, "range_ms": [ 731.859543826431, 743.015272077173 ] }, "cold": { "count": 10, "median_ms": 772.390644531697, "p95_ms": 781.3249886967242, "range_ms": [ 768.2700520381331, 781.3249886967242 ] }, "warm_median_decisions_per_second": 86.90719475239233 } }, "performance": { "1": { "warm": { "count": 10, "median_ms": 23.32297363318503, "p95_ms": 34.777252934873104, "range_ms": [ 23.126515094190836, 34.777252934873104 ] }, "cold": { "count": 10, "median_ms": 181.3432709313929, "p95_ms": 193.60135402530432, "range_ms": [ 178.26369171962142, 193.60135402530432 ] }, "warm_median_decisions_per_second": 42.87617932977263 }, "8": { "warm": { "count": 10, "median_ms": 100.68796202540398, "p95_ms": 101.85512574389577, "range_ms": [ 96.88412584364414, 101.85512574389577 ] }, "cold": { "count": 10, "median_ms": 256.44394662231207, "p95_ms": 266.86199894174933, "range_ms": [ 254.43977303802967, 266.86199894174933 ] }, "warm_median_decisions_per_second": 79.45339084310365 }, "16": { "warm": { "count": 10, "median_ms": 198.03814101032913, "p95_ms": 202.5389182381332, "range_ms": [ 195.7407076843083, 202.5389182381332 ] }, "cold": { "count": 10, "median_ms": 357.35773434862494, "p95_ms": 371.45887641236186, "range_ms": [ 354.1008699685335, 371.45887641236186 ] }, "warm_median_decisions_per_second": 80.79251763510284 }, "32": { "warm": { "count": 10, "median_ms": 391.90001506358385, "p95_ms": 398.1366469524801, "range_ms": [ 384.79119492694736, 398.1366469524801 ] }, "cold": { "count": 10, "median_ms": 549.7453005518764, "p95_ms": 553.3181391656399, "range_ms": [ 545.2003879472613, 553.3181391656399 ] }, "warm_median_decisions_per_second": 81.65347989284501 }, "64": { "warm": { "count": 10, "median_ms": 788.8812434393913, "p95_ms": 793.8674939796329, "range_ms": [ 778.851400129497, 793.8674939796329 ] }, "cold": { "count": 10, "median_ms": 947.814577491954, "p95_ms": 950.3754759207368, "range_ms": [ 934.6124189905822, 950.3754759207368 ] }, "warm_median_decisions_per_second": 81.12754680409262 } } }, "artifact": { "sha256": "b710efc4c0d048ee61eed92c5fef5ce323a4d17e7c51f9f0533cc72ae50818ea", "bytes": 12669646592 }, "serving_configuration": "Winnow llama.cpp runtime; ctx=8192; decision-context=8192; decision-parallel=4; chat-parallel=1; KV=Q8_0; selected head; optimized pipeline; memory=auto; full GPU offload; no vision projector", "source_pins": { "kev_git": "bd058057ad0aa9df3dd6d14e3542c95a5ce367b2", "laya_git": "42626c348753fbb17572a813127df2278a1ec527", "jevbench_git": "c6004e008ffba24aec091261ca1a5c02f7324702", "kev_4b": "f780095d6511e6304868ff46ca2b74ee3eba29d8", "kev_9b": "6281032426a9ca3a08374d137bf9dfd8afd78e8a", "laya": "1c5edc17a7acd8701df6fc341c0d179f1c62c982", "laya_typed": "f9ab0b228f0fc0f14d873dbc99038f135c2da1b2", "typed_decisions": "ea9306458d6e9563628369a3d1e72e362fb381d2" }, "dataset_manifest_sha256": { "kev_v9": "3c4f0be94509a3612678bfd3a30fd99a8d0ca3c47ddfe7318075d95b2fa365e4", "jevbench": "0983ad3e73a34e7c55b02b5d09ed983aeccabef0253bb31be43a2e6230d58f7b" }, "inference_build": { "server_sha256": "ab75f69c68ccc8063d44aec6ac2a24ce07e5698f8175a263973d46c029256e0f", "runtime_lock_sha256": "21e6e69d8e43e318089e77459d102f05ee2cad266999c7860e64048f69c89e84", "cuda_arch": "120", "compiler_cache_sha256": "c105481bdf985eb42f79dee77c13caf57c36e2e419acafad975c6c04b9c81adf" } }, { "model": "Winnow-12B BF16", "precision": "BF16", "quality_percent": { "jevbench": 85.28138528138528, "kev_v9": 81.45315487571702, "kev_additions": 69.23076923076923, "typed": 70.19999999999999 }, "counts": { "jevbench": 231, "kev_v9": 1046, "kev_additions": 390, "typed": 2000 }, "calibration": { "jev_brier": 0.20424101171755193, "jev_ece": 0.06547912754284907, "kev_brier": 0.28568609930480093, "kev_ece": 0.09889472147602915, "typed_teacher_brier_non_score": 0.22675176129911292, "typed_ece": 0.15511670583963688 }, "memory_gib": { "gpu": "NVIDIA RTX PRO 5000 Blackwell", "device_baseline_mib": 2.0, "load_vram_gib": 23.220703125, "idle_vram_gib": 24.771484375, "inference_vram_gib": 24.796875, "load_ram_pss_gib": 22.580784797668457, "idle_ram_pss_gib": 0.6659975051879883, "inference_ram_pss_gib": 0.7166461944580078 }, "performance": { "performance-short": { "1": { "warm": { "count": 10, "median_ms": 33.53214240632951, "p95_ms": 43.883658945560455, "range_ms": [ 33.169500064104795, 43.883658945560455 ] }, "cold": { "count": 10, "median_ms": 74.17859323322773, "p95_ms": 88.81977200508118, "range_ms": [ 73.65309121087193, 88.81977200508118 ] }, "warm_median_decisions_per_second": 29.822132683393367 }, "8": { "warm": { "count": 10, "median_ms": 107.99862560816109, "p95_ms": 110.18604878336191, "range_ms": [ 107.16228932142258, 110.18604878336191 ] }, "cold": { "count": 10, "median_ms": 147.56763633340597, "p95_ms": 153.14696403220296, "range_ms": [ 146.64038131013513, 153.14696403220296 ] }, "warm_median_decisions_per_second": 74.07501674165256 }, "16": { "warm": { "count": 10, "median_ms": 216.6108468081802, "p95_ms": 222.67627017572522, "range_ms": [ 215.47124814242125, 222.67627017572522 ] }, "cold": { "count": 10, "median_ms": 258.43513989821076, "p95_ms": 268.3673081919551, "range_ms": [ 257.0353946648538, 268.3673081919551 ] }, "warm_median_decisions_per_second": 73.86518374201641 }, "32": { "warm": { "count": 10, "median_ms": 427.29172739200294, "p95_ms": 429.2222601361573, "range_ms": [ 426.3705341145396, 429.2222601361573 ] }, "cold": { "count": 10, "median_ms": 467.9329860955477, "p95_ms": 476.0954240337014, "range_ms": [ 466.92013274878263, 476.0954240337014 ] }, "warm_median_decisions_per_second": 74.89028677272468 }, "64": { "warm": { "count": 10, "median_ms": 856.0223490931094, "p95_ms": 857.3235273361206, "range_ms": [ 854.7417223453522, 857.3235273361206 ] }, "cold": { "count": 10, "median_ms": 897.3894559312612, "p95_ms": 901.5537169761956, "range_ms": [ 895.5069477669895, 901.5537169761956 ] }, "warm_median_decisions_per_second": 74.764403134805 } }, "performance": { "1": { "warm": { "count": 10, "median_ms": 34.854388097301126, "p95_ms": 48.964081797748804, "range_ms": [ 34.537527710199356, 48.964081797748804 ] }, "cold": { "count": 10, "median_ms": 195.44652244076133, "p95_ms": 207.5622510164976, "range_ms": [ 193.3445748873055, 207.5622510164976 ] }, "warm_median_decisions_per_second": 28.69079202332727 }, "8": { "warm": { "count": 10, "median_ms": 112.90686530992389, "p95_ms": 116.05708813294768, "range_ms": [ 112.2279530391097, 116.05708813294768 ] }, "cold": { "count": 10, "median_ms": 273.41523696668446, "p95_ms": 277.3972353897989, "range_ms": [ 272.6574372500181, 277.3972353897989 ] }, "warm_median_decisions_per_second": 70.85485880810158 }, "16": { "warm": { "count": 10, "median_ms": 226.20371961966157, "p95_ms": 232.85513184964657, "range_ms": [ 224.97101593762636, 232.85513184964657 ] }, "cold": { "count": 10, "median_ms": 387.9560462664813, "p95_ms": 392.6410130225122, "range_ms": [ 387.3477568849921, 392.6410130225122 ] }, "warm_median_decisions_per_second": 70.73270071288998 }, "32": { "warm": { "count": 10, "median_ms": 446.2287644855678, "p95_ms": 449.11305233836174, "range_ms": [ 445.53177896887064, 449.11305233836174 ] }, "cold": { "count": 10, "median_ms": 607.0701309945434, "p95_ms": 612.5219059176743, "range_ms": [ 606.2590656802058, 612.5219059176743 ] }, "warm_median_decisions_per_second": 71.7120960072823 }, "64": { "warm": { "count": 10, "median_ms": 893.4033433906734, "p95_ms": 905.7790436781943, "range_ms": [ 892.6470330916345, 905.7790436781943 ] }, "cold": { "count": 10, "median_ms": 1054.341682465747, "p95_ms": 1057.7269801869988, "range_ms": [ 1053.0990920960903, 1057.7269801869988 ] }, "warm_median_decisions_per_second": 71.63617695575788 } } }, "artifact": { "sha256": "a7ec058c65b4e73da9f6066a0239b449c74d644c234a5f901aebb19e56de3a22", "bytes": 23832065792 }, "serving_configuration": "Winnow llama.cpp runtime; ctx=8192; decision-context=8192; decision-parallel=4; chat-parallel=1; KV=Q8_0; selected head; optimized pipeline; memory=auto; full GPU offload; no vision projector", "source_pins": { "kev_git": "bd058057ad0aa9df3dd6d14e3542c95a5ce367b2", "laya_git": "42626c348753fbb17572a813127df2278a1ec527", "jevbench_git": "c6004e008ffba24aec091261ca1a5c02f7324702", "kev_4b": "f780095d6511e6304868ff46ca2b74ee3eba29d8", "kev_9b": "6281032426a9ca3a08374d137bf9dfd8afd78e8a", "laya": "1c5edc17a7acd8701df6fc341c0d179f1c62c982", "laya_typed": "f9ab0b228f0fc0f14d873dbc99038f135c2da1b2", "typed_decisions": "ea9306458d6e9563628369a3d1e72e362fb381d2" }, "dataset_manifest_sha256": { "kev_v9": "3c4f0be94509a3612678bfd3a30fd99a8d0ca3c47ddfe7318075d95b2fa365e4", "jevbench": "0983ad3e73a34e7c55b02b5d09ed983aeccabef0253bb31be43a2e6230d58f7b" }, "inference_build": { "server_sha256": "ab75f69c68ccc8063d44aec6ac2a24ce07e5698f8175a263973d46c029256e0f", "runtime_lock_sha256": "21e6e69d8e43e318089e77459d102f05ee2cad266999c7860e64048f69c89e84", "cuda_arch": "120", "compiler_cache_sha256": "c105481bdf985eb42f79dee77c13caf57c36e2e419acafad975c6c04b9c81adf" } }, { "model": "Stock Gemma 4 12B IT Q8", "precision": "Q8_0", "quality_percent": { "jevbench": 83.54978354978356, "kev_v9": 78.68068833652008, "kev_additions": 66.66666666666666, "typed": 71.3 }, "counts": { "jevbench": 231, "kev_v9": 1046, "kev_additions": 390, "typed": 2000 }, "calibration": { "jev_brier": 0.27034976841267755, "jev_ece": 0.1316843789806068, "kev_brier": 0.3702287880491706, "kev_ece": 0.16323469713749836, "typed_teacher_brier_non_score": 0.316792420641541, "typed_ece": 0.24805315512618353 }, "memory_gib": { "gpu": "NVIDIA RTX PRO 5000 Blackwell", "device_baseline_mib": 2.0, "load_vram_gib": 12.82421875, "idle_vram_gib": 14.333984375, "inference_vram_gib": 14.35546875, "load_ram_pss_gib": 12.191630363464355, "idle_ram_pss_gib": 0.5481204986572266, "inference_ram_pss_gib": 0.5979843139648438 }, "performance": { "performance-short": { "1": { "warm": { "count": 10, "median_ms": 22.46146695688367, "p95_ms": 30.770714860409498, "range_ms": [ 22.075897082686424, 30.770714860409498 ] }, "cold": { "count": 10, "median_ms": 52.51622456125915, "p95_ms": 60.364149045199156, "range_ms": [ 51.94238806143403, 60.364149045199156 ] }, "warm_median_decisions_per_second": 44.520689673544865 }, "8": { "warm": { "count": 10, "median_ms": 92.13604708202183, "p95_ms": 93.77083322033286, "range_ms": [ 91.1752269603312, 93.77083322033286 ] }, "cold": { "count": 10, "median_ms": 120.54571812041104, "p95_ms": 124.4972855783999, "range_ms": [ 120.1186771504581, 124.4972855783999 ] }, "warm_median_decisions_per_second": 86.82812268772719 }, "16": { "warm": { "count": 10, "median_ms": 185.5551926419139, "p95_ms": 188.82568273693323, "range_ms": [ 184.85999898985028, 188.82568273693323 ] }, "cold": { "count": 10, "median_ms": 217.12910407222807, "p95_ms": 222.90433011949062, "range_ms": [ 216.16374514997005, 222.90433011949062 ] }, "warm_median_decisions_per_second": 86.22771355624063 }, "32": { "warm": { "count": 10, "median_ms": 362.9481350071728, "p95_ms": 364.25275122746825, "range_ms": [ 362.4884523451328, 364.25275122746825 ] }, "cold": { "count": 10, "median_ms": 393.8138112425804, "p95_ms": 397.9031350463629, "range_ms": [ 393.2893369346857, 397.9031350463629 ] }, "warm_median_decisions_per_second": 88.16686714581849 }, "64": { "warm": { "count": 10, "median_ms": 727.991261985153, "p95_ms": 729.7718632034957, "range_ms": [ 726.691781077534, 729.7718632034957 ] }, "cold": { "count": 10, "median_ms": 759.8112525884062, "p95_ms": 765.421356074512, "range_ms": [ 757.9216249287128, 765.421356074512 ] }, "warm_median_decisions_per_second": 87.91314311311783 } }, "performance": { "1": { "warm": { "count": 10, "median_ms": 23.851238656789064, "p95_ms": 31.665181275457144, "range_ms": [ 23.5066469758749, 31.665181275457144 ] }, "cold": { "count": 10, "median_ms": 178.94592159427702, "p95_ms": 190.3981356881559, "range_ms": [ 177.06820787861943, 190.3981356881559 ] }, "warm_median_decisions_per_second": 41.926543706582635 }, "8": { "warm": { "count": 10, "median_ms": 97.29194873943925, "p95_ms": 98.95601915195584, "range_ms": [ 96.71131893992424, 98.95601915195584 ] }, "cold": { "count": 10, "median_ms": 253.41987097635865, "p95_ms": 259.4043822027743, "range_ms": [ 251.87801895663142, 259.4043822027743 ] }, "warm_median_decisions_per_second": 82.22674233224645 }, "16": { "warm": { "count": 10, "median_ms": 194.9084191583097, "p95_ms": 199.53563576564193, "range_ms": [ 194.04081907123327, 199.53563576564193 ] }, "cold": { "count": 10, "median_ms": 352.85464976914227, "p95_ms": 360.91970512643456, "range_ms": [ 351.35736502707005, 360.91970512643456 ] }, "warm_median_decisions_per_second": 82.08983515999061 }, "32": { "warm": { "count": 10, "median_ms": 382.41670886054635, "p95_ms": 383.51830979809165, "range_ms": [ 381.01493706926703, 383.51830979809165 ] }, "cold": { "count": 10, "median_ms": 539.1273065470159, "p95_ms": 545.6521101295948, "range_ms": [ 538.4233240038157, 545.6521101295948 ] }, "warm_median_decisions_per_second": 83.67835206612077 }, "64": { "warm": { "count": 10, "median_ms": 766.0538714844733, "p95_ms": 768.1685579009354, "range_ms": [ 764.6735701709986, 768.1685579009354 ] }, "cold": { "count": 10, "median_ms": 923.574757296592, "p95_ms": 928.5005992278457, "range_ms": [ 922.1836552023888, 928.5005992278457 ] }, "warm_median_decisions_per_second": 83.54503825687823 } } }, "artifact": { "sha256": "f20e7ff1be28c283eeeb18fc895733791c56a5851d5cd3fe9691b7f7d12afa72", "bytes": 12669647680 }, "serving_configuration": "Winnow llama.cpp runtime; ctx=8192; decision-context=8192; decision-parallel=4; chat-parallel=1; KV=Q8_0; selected head; optimized pipeline; memory=auto; full GPU offload; no vision projector", "source_pins": { "kev_git": "bd058057ad0aa9df3dd6d14e3542c95a5ce367b2", "laya_git": "42626c348753fbb17572a813127df2278a1ec527", "jevbench_git": "c6004e008ffba24aec091261ca1a5c02f7324702", "kev_4b": "f780095d6511e6304868ff46ca2b74ee3eba29d8", "kev_9b": "6281032426a9ca3a08374d137bf9dfd8afd78e8a", "laya": "1c5edc17a7acd8701df6fc341c0d179f1c62c982", "laya_typed": "f9ab0b228f0fc0f14d873dbc99038f135c2da1b2", "typed_decisions": "ea9306458d6e9563628369a3d1e72e362fb381d2" }, "dataset_manifest_sha256": { "kev_v9": "3c4f0be94509a3612678bfd3a30fd99a8d0ca3c47ddfe7318075d95b2fa365e4", "jevbench": "0983ad3e73a34e7c55b02b5d09ed983aeccabef0253bb31be43a2e6230d58f7b" }, "inference_build": { "server_sha256": "ab75f69c68ccc8063d44aec6ac2a24ce07e5698f8175a263973d46c029256e0f", "runtime_lock_sha256": "21e6e69d8e43e318089e77459d102f05ee2cad266999c7860e64048f69c89e84", "cuda_arch": "120", "compiler_cache_sha256": "c105481bdf985eb42f79dee77c13caf57c36e2e419acafad975c6c04b9c81adf" } }, { "model": "Stock Gemma 4 12B IT BF16", "precision": "BF16", "quality_percent": { "jevbench": 83.54978354978356, "kev_v9": 78.29827915869981, "kev_additions": 65.8974358974359, "typed": 71.39999999999999 }, "counts": { "jevbench": 231, "kev_v9": 1046, "kev_additions": 390, "typed": 2000 }, "calibration": { "jev_brier": 0.26935175054650906, "jev_ece": 0.13191123135721156, "kev_brier": 0.37384468803633186, "kev_ece": 0.17031457341371484, "typed_teacher_brier_non_score": 0.3173045037688962, "typed_ece": 0.2469585567812167 }, "memory_gib": { "gpu": "NVIDIA RTX PRO 5000 Blackwell", "device_baseline_mib": 2.0, "load_vram_gib": 23.220703125, "idle_vram_gib": 24.771484375, "inference_vram_gib": 24.796875, "load_ram_pss_gib": 22.587621688842773, "idle_ram_pss_gib": 0.6732587814331055, "inference_ram_pss_gib": 0.7233238220214844 }, "performance": { "performance-short": { "1": { "warm": { "count": 10, "median_ms": 34.141618525609374, "p95_ms": 41.745335794985294, "range_ms": [ 33.73613115400076, 41.745335794985294 ] }, "cold": { "count": 10, "median_ms": 75.02603204920888, "p95_ms": 89.06905213370919, "range_ms": [ 74.07003827393055, 89.06905213370919 ] }, "warm_median_decisions_per_second": 29.28976548812141 }, "8": { "warm": { "count": 10, "median_ms": 108.13127923756838, "p95_ms": 108.66840602830052, "range_ms": [ 107.68243437632918, 108.66840602830052 ] }, "cold": { "count": 10, "median_ms": 148.21077208034694, "p95_ms": 152.73841423913836, "range_ms": [ 147.22343580797315, 152.73841423913836 ] }, "warm_median_decisions_per_second": 73.9841427606133 }, "16": { "warm": { "count": 10, "median_ms": 216.71691467054188, "p95_ms": 223.66143111139536, "range_ms": [ 215.90717369690537, 223.66143111139536 ] }, "cold": { "count": 10, "median_ms": 258.0273279454559, "p95_ms": 268.22592318058014, "range_ms": [ 257.2539518587291, 268.22592318058014 ] }, "warm_median_decisions_per_second": 73.82903187009455 }, "32": { "warm": { "count": 10, "median_ms": 427.47937608510256, "p95_ms": 428.38083393871784, "range_ms": [ 426.6806561499834, 428.38083393871784 ] }, "cold": { "count": 10, "median_ms": 467.7976402454078, "p95_ms": 472.7875650860369, "range_ms": [ 467.0621966943145, 472.7875650860369 ] }, "warm_median_decisions_per_second": 74.85741252141587 }, "64": { "warm": { "count": 10, "median_ms": 855.5614438373595, "p95_ms": 856.8025063723326, "range_ms": [ 854.8735701479018, 856.8025063723326 ] }, "cold": { "count": 10, "median_ms": 896.0815852042288, "p95_ms": 901.9729560241103, "range_ms": [ 895.2084640040994, 901.9729560241103 ] }, "warm_median_decisions_per_second": 74.80467996891906 } }, "performance": { "1": { "warm": { "count": 10, "median_ms": 34.927206579595804, "p95_ms": 49.98836200684309, "range_ms": [ 34.57708191126585, 49.98836200684309 ] }, "cold": { "count": 10, "median_ms": 194.9117931071669, "p95_ms": 205.73526807129383, "range_ms": [ 192.28755496442318, 205.73526807129383 ] }, "warm_median_decisions_per_second": 28.630975618422116 }, "8": { "warm": { "count": 10, "median_ms": 112.50834492966533, "p95_ms": 114.36860868707299, "range_ms": [ 111.96773778647184, 114.36860868707299 ] }, "cold": { "count": 10, "median_ms": 273.30744452774525, "p95_ms": 277.10055792704225, "range_ms": [ 272.5763861089945, 277.10055792704225 ] }, "warm_median_decisions_per_second": 71.10583668261413 }, "16": { "warm": { "count": 10, "median_ms": 226.3854960910976, "p95_ms": 231.87217162922025, "range_ms": [ 225.70885298773646, 231.87217162922025 ] }, "cold": { "count": 10, "median_ms": 388.3433099836111, "p95_ms": 392.61183608323336, "range_ms": [ 387.42467015981674, 392.61183608323336 ] }, "warm_median_decisions_per_second": 70.67590581669417 }, "32": { "warm": { "count": 10, "median_ms": 446.0229475516826, "p95_ms": 446.9721997156739, "range_ms": [ 444.9301161803305, 446.9721997156739 ] }, "cold": { "count": 10, "median_ms": 606.8045513238758, "p95_ms": 611.3648116588593, "range_ms": [ 606.0819658450782, 611.3648116588593 ] }, "warm_median_decisions_per_second": 71.74518749686533 }, "64": { "warm": { "count": 10, "median_ms": 891.7435111943632, "p95_ms": 901.2053925544024, "range_ms": [ 890.0642930530012, 901.2053925544024 ] }, "cold": { "count": 10, "median_ms": 1054.1123251896352, "p95_ms": 1057.0517675951123, "range_ms": [ 1053.6570539698005, 1057.0517675951123 ] }, "warm_median_decisions_per_second": 71.76951578182063 } } }, "artifact": { "sha256": "5a5eefea73350705c6753105b725a51301d653e8d6646173db29a7e8da8e6efd", "bytes": 23832066656 }, "serving_configuration": "Winnow llama.cpp runtime; ctx=8192; decision-context=8192; decision-parallel=4; chat-parallel=1; KV=Q8_0; selected head; optimized pipeline; memory=auto; full GPU offload; no vision projector", "source_pins": { "kev_git": "bd058057ad0aa9df3dd6d14e3542c95a5ce367b2", "laya_git": "42626c348753fbb17572a813127df2278a1ec527", "jevbench_git": "c6004e008ffba24aec091261ca1a5c02f7324702", "kev_4b": "f780095d6511e6304868ff46ca2b74ee3eba29d8", "kev_9b": "6281032426a9ca3a08374d137bf9dfd8afd78e8a", "laya": "1c5edc17a7acd8701df6fc341c0d179f1c62c982", "laya_typed": "f9ab0b228f0fc0f14d873dbc99038f135c2da1b2", "typed_decisions": "ea9306458d6e9563628369a3d1e72e362fb381d2" }, "dataset_manifest_sha256": { "kev_v9": "3c4f0be94509a3612678bfd3a30fd99a8d0ca3c47ddfe7318075d95b2fa365e4", "jevbench": "0983ad3e73a34e7c55b02b5d09ed983aeccabef0253bb31be43a2e6230d58f7b" }, "inference_build": { "server_sha256": "ab75f69c68ccc8063d44aec6ac2a24ce07e5698f8175a263973d46c029256e0f", "runtime_lock_sha256": "21e6e69d8e43e318089e77459d102f05ee2cad266999c7860e64048f69c89e84", "cuda_arch": "120", "compiler_cache_sha256": "c105481bdf985eb42f79dee77c13caf57c36e2e419acafad975c6c04b9c81adf" } }, { "model": "Kev-9B BF16", "precision": "BF16", "quality_percent": { "jevbench": 76.19047619047619, "kev_v9": 78.20267686424475, "kev_additions": 68.97435897435898, "typed": 71.6 }, "counts": { "jevbench": 231, "kev_v9": 1046, "kev_additions": 390, "typed": 2000 }, "calibration": { "jev_brier": 0.3540905585960517, "jev_ece": 0.16230247700094697, "kev_brier": 0.31815661474414736, "kev_ece": 0.100808307899414, "typed_teacher_brier_non_score": 0.15618699632615016, "typed_ece": 0.059057252225222456 }, "memory_gib": { "gpu": "NVIDIA RTX PRO 5000 Blackwell", "device_baseline_mib": 2.0, "load_vram_gib": 32.29296875, "idle_vram_gib": 15.23828125, "inference_vram_gib": 21.8671875, "load_ram_pss_gib": 47.3271598815918, "idle_ram_pss_gib": 1.6994895935058594, "inference_ram_pss_gib": 2.419340133666992 }, "performance": { "performance-short": { "1": { "warm": { "count": 10, "median_ms": 66.0799175966531, "p95_ms": 68.83646594360471, "range_ms": [ 64.38081990927458, 68.83646594360471 ] }, "cold": { "count": 10, "median_ms": 65.56061049923301, "p95_ms": 69.4022267125547, "range_ms": [ 64.67744009569287, 69.4022267125547 ] }, "warm_median_decisions_per_second": 15.133190784285258 }, "8": { "warm": { "count": 10, "median_ms": 72.50639772973955, "p95_ms": 77.58744899183512, "range_ms": [ 72.30309210717678, 77.58744899183512 ] }, "cold": { "count": 10, "median_ms": 72.83194921910763, "p95_ms": 76.30407717078924, "range_ms": [ 72.55592988803983, 76.30407717078924 ] }, "warm_median_decisions_per_second": 110.3350911159483 }, "16": { "warm": { "count": 10, "median_ms": 117.60589014738798, "p95_ms": 120.17511110752821, "range_ms": [ 115.76627707108855, 120.17511110752821 ] }, "cold": { "count": 10, "median_ms": 119.11514005623758, "p95_ms": 120.19340693950653, "range_ms": [ 118.12467919662595, 120.19340693950653 ] }, "warm_median_decisions_per_second": 136.04760764914255 }, "32": { "warm": { "count": 10, "median_ms": 229.1738234926015, "p95_ms": 231.51671793311834, "range_ms": [ 228.54166710749269, 231.51671793311834 ] }, "cold": { "count": 10, "median_ms": 231.84350552037358, "p95_ms": 234.30478060618043, "range_ms": [ 230.4078508168459, 234.30478060618043 ] }, "warm_median_decisions_per_second": 139.6320029588068 }, "64": { "warm": { "count": 10, "median_ms": 484.2144048307091, "p95_ms": 486.98815423995256, "range_ms": [ 483.16515097394586, 486.98815423995256 ] }, "cold": { "count": 10, "median_ms": 494.19798236340284, "p95_ms": 494.99255511909723, "range_ms": [ 493.5324052348733, 494.99255511909723 ] }, "warm_median_decisions_per_second": 132.17285434202574 } }, "performance": { "1": { "warm": { "count": 10, "median_ms": 71.03875442408025, "p95_ms": 75.37550991401076, "range_ms": [ 70.50783606246114, 75.37550991401076 ] }, "cold": { "count": 10, "median_ms": 151.8427466508001, "p95_ms": 155.29751405119896, "range_ms": [ 151.14088403061032, 155.29751405119896 ] }, "warm_median_decisions_per_second": 14.076823391782705 }, "8": { "warm": { "count": 10, "median_ms": 77.25602923892438, "p95_ms": 81.38543087989092, "range_ms": [ 76.64880808442831, 81.38543087989092 ] }, "cold": { "count": 10, "median_ms": 158.6230390239507, "p95_ms": 163.41657424345613, "range_ms": [ 157.5089693069458, 163.41657424345613 ] }, "warm_median_decisions_per_second": 103.55178849871966 }, "16": { "warm": { "count": 10, "median_ms": 82.72549859248102, "p95_ms": 87.11598766967654, "range_ms": [ 82.24265603348613, 87.11598766967654 ] }, "cold": { "count": 10, "median_ms": 164.65989546850324, "p95_ms": 169.00390107184649, "range_ms": [ 164.0263837762177, 169.00390107184649 ] }, "warm_median_decisions_per_second": 193.41074121316026 }, "32": { "warm": { "count": 10, "median_ms": 135.10010042227805, "p95_ms": 137.17515813186765, "range_ms": [ 134.41574899479747, 137.17515813186765 ] }, "cold": { "count": 10, "median_ms": 219.00241915136576, "p95_ms": 221.23930510133505, "range_ms": [ 217.83947385847569, 221.23930510133505 ] }, "warm_median_decisions_per_second": 236.86140794846656 }, "64": { "warm": { "count": 10, "median_ms": 231.3950948882848, "p95_ms": 235.025763977319, "range_ms": [ 230.0196811556816, 235.025763977319 ] }, "cold": { "count": 10, "median_ms": 316.5164175443351, "p95_ms": 320.0176330283284, "range_ms": [ 314.9941503070295, 320.0176330283284 ] }, "warm_median_decisions_per_second": 276.5832181140164 } } }, "artifact": {}, "serving_configuration": "KEV_DTYPE=bf16 KEV_MERGE=1 KEV_PREFIX_CACHE=1; unused allocator cache released after load", "source_pins": { "kev_git": "bd058057ad0aa9df3dd6d14e3542c95a5ce367b2", "laya_git": "42626c348753fbb17572a813127df2278a1ec527", "jevbench_git": "c6004e008ffba24aec091261ca1a5c02f7324702", "kev_4b": "f780095d6511e6304868ff46ca2b74ee3eba29d8", "kev_9b": "6281032426a9ca3a08374d137bf9dfd8afd78e8a", "laya": "1c5edc17a7acd8701df6fc341c0d179f1c62c982", "laya_typed": "f9ab0b228f0fc0f14d873dbc99038f135c2da1b2", "typed_decisions": "ea9306458d6e9563628369a3d1e72e362fb381d2" }, "dataset_manifest_sha256": { "kev_v9": "3c4f0be94509a3612678bfd3a30fd99a8d0ca3c47ddfe7318075d95b2fa365e4", "jevbench": "0983ad3e73a34e7c55b02b5d09ed983aeccabef0253bb31be43a2e6230d58f7b" }, "inference_build": {} }, { "model": "Kev-4B BF16", "precision": "BF16", "quality_percent": { "jevbench": 71.42857142857143, "kev_v9": 75.62141491395793, "kev_additions": 63.589743589743584, "typed": 65.75 }, "counts": { "jevbench": 231, "kev_v9": 1046, "kev_additions": 390, "typed": 2000 }, "calibration": { "jev_brier": 0.45430566532075206, "jev_ece": 0.19253563884526956, "kev_brier": 0.357785537147028, "kev_ece": 0.11677672947493024, "typed_teacher_brier_non_score": 0.22585355939972085, "typed_ece": 0.1253794714471447 }, "memory_gib": { "gpu": "NVIDIA RTX PRO 5000 Blackwell", "device_baseline_mib": 2.0, "load_vram_gib": 17.57421875, "idle_vram_gib": 8.28125, "inference_vram_gib": 14.892578125, "load_ram_pss_gib": 23.110424041748047, "idle_ram_pss_gib": 1.6300640106201172, "inference_ram_pss_gib": 2.4912405014038086 }, "performance": { "performance-short": { "1": { "warm": { "count": 10, "median_ms": 65.64339064061642, "p95_ms": 67.66112381592393, "range_ms": [ 65.52932038903236, 67.66112381592393 ] }, "cold": { "count": 10, "median_ms": 65.76349004171789, "p95_ms": 67.32068909332156, "range_ms": [ 65.40474807843566, 67.32068909332156 ] }, "warm_median_decisions_per_second": 15.233826136050574 }, "8": { "warm": { "count": 10, "median_ms": 72.70564697682858, "p95_ms": 79.85566509887576, "range_ms": [ 72.43038015440106, 79.85566509887576 ] }, "cold": { "count": 10, "median_ms": 72.54690118134022, "p95_ms": 74.15724685415626, "range_ms": [ 72.02649116516113, 74.15724685415626 ] }, "warm_median_decisions_per_second": 110.03271867657838 }, "16": { "warm": { "count": 10, "median_ms": 77.47087301686406, "p95_ms": 78.55709921568632, "range_ms": [ 77.14755507186055, 78.55709921568632 ] }, "cold": { "count": 10, "median_ms": 77.62656570412219, "p95_ms": 78.94049910828471, "range_ms": [ 77.46943598613143, 78.94049910828471 ] }, "warm_median_decisions_per_second": 206.52923320635716 }, "32": { "warm": { "count": 10, "median_ms": 148.88300257734954, "p95_ms": 153.7636579014361, "range_ms": [ 148.60826171934605, 153.7636579014361 ] }, "cold": { "count": 10, "median_ms": 150.5928838159889, "p95_ms": 151.1529739946127, "range_ms": [ 150.2681067213416, 151.1529739946127 ] }, "warm_median_decisions_per_second": 214.9338705294781 }, "64": { "warm": { "count": 10, "median_ms": 312.89613409899175, "p95_ms": 314.574324991554, "range_ms": [ 312.318773008883, 314.574324991554 ] }, "cold": { "count": 10, "median_ms": 319.24319406971335, "p95_ms": 320.00197703018785, "range_ms": [ 318.71265126392245, 320.00197703018785 ] }, "warm_median_decisions_per_second": 204.54071823000587 } }, "performance": { "1": { "warm": { "count": 10, "median_ms": 71.24498765915632, "p95_ms": 74.90490330383182, "range_ms": [ 70.84893900901079, 74.90490330383182 ] }, "cold": { "count": 10, "median_ms": 133.07837629690766, "p95_ms": 134.00867069140077, "range_ms": [ 132.59583385661244, 134.00867069140077 ] }, "warm_median_decisions_per_second": 14.036075138143156 }, "8": { "warm": { "count": 10, "median_ms": 78.0957960523665, "p95_ms": 79.34773107990623, "range_ms": [ 77.41492474451661, 79.34773107990623 ] }, "cold": { "count": 10, "median_ms": 139.4394184462726, "p95_ms": 142.5566808320582, "range_ms": [ 138.79002584144473, 142.5566808320582 ] }, "warm_median_decisions_per_second": 102.43829251238652 }, "16": { "warm": { "count": 10, "median_ms": 83.62268563359976, "p95_ms": 84.26053682342172, "range_ms": [ 83.19316618144512, 84.26053682342172 ] }, "cold": { "count": 10, "median_ms": 145.1625218614936, "p95_ms": 147.97593792900443, "range_ms": [ 144.65159736573696, 147.97593792900443 ] }, "warm_median_decisions_per_second": 191.33563911239858 }, "32": { "warm": { "count": 10, "median_ms": 102.64253127388656, "p95_ms": 103.2784916460514, "range_ms": [ 102.27458225563169, 103.2784916460514 ] }, "cold": { "count": 10, "median_ms": 164.34098314493895, "p95_ms": 165.785264223814, "range_ms": [ 163.7231931090355, 165.785264223814 ] }, "warm_median_decisions_per_second": 311.7616021628762 }, "64": { "warm": { "count": 10, "median_ms": 179.33518323116004, "p95_ms": 180.0042991526425, "range_ms": [ 178.8237178698182, 180.0042991526425 ] }, "cold": { "count": 10, "median_ms": 240.78315845690668, "p95_ms": 242.358619812876, "range_ms": [ 240.2559290640056, 242.358619812876 ] }, "warm_median_decisions_per_second": 356.8736421202139 } } }, "artifact": {}, "serving_configuration": "KEV_DTYPE=bf16 KEV_MERGE=1 KEV_PREFIX_CACHE=1; unused allocator cache released after load", "source_pins": { "kev_git": "bd058057ad0aa9df3dd6d14e3542c95a5ce367b2", "laya_git": "42626c348753fbb17572a813127df2278a1ec527", "jevbench_git": "c6004e008ffba24aec091261ca1a5c02f7324702", "kev_4b": "f780095d6511e6304868ff46ca2b74ee3eba29d8", "kev_9b": "6281032426a9ca3a08374d137bf9dfd8afd78e8a", "laya": "1c5edc17a7acd8701df6fc341c0d179f1c62c982", "laya_typed": "f9ab0b228f0fc0f14d873dbc99038f135c2da1b2", "typed_decisions": "ea9306458d6e9563628369a3d1e72e362fb381d2" }, "dataset_manifest_sha256": { "kev_v9": "3c4f0be94509a3612678bfd3a30fd99a8d0ca3c47ddfe7318075d95b2fa365e4", "jevbench": "0983ad3e73a34e7c55b02b5d09ed983aeccabef0253bb31be43a2e6230d58f7b" }, "inference_build": {} }, { "model": "Laya Typed Decisions", "precision": "float32 weights; bfloat16 autocast", "quality_percent": { "jevbench": 53.67965367965368, "kev_v9": 55.64053537284895, "kev_additions": 36.666666666666664, "typed": 76.75 }, "counts": { "jevbench": 231, "kev_v9": 1046, "kev_additions": 390, "typed": 2000 }, "calibration": { "jev_brier": 0.5087844836796537, "jev_ece": 0.07119307359307356, "kev_brier": 0.5406436657135641, "kev_ece": 0.051804615299576945, "typed_teacher_brier_non_score": 0.06574973716973499, "typed_ece": 0.2146720270839175 }, "memory_gib": { "gpu": "NVIDIA RTX PRO 5000 Blackwell", "device_baseline_mib": 2.0, "load_vram_gib": 1.822265625, "idle_vram_gib": 1.919921875, "inference_vram_gib": 4.94140625, "load_ram_pss_gib": 3.2509498596191406, "idle_ram_pss_gib": 1.0975418090820312, "inference_ram_pss_gib": 2.0477752685546875 }, "performance": { "performance-short": { "1": { "warm": { "count": 10, "median_ms": 22.845126455649734, "p95_ms": 24.33396875858307, "range_ms": [ 22.740793880075216, 24.33396875858307 ] }, "cold": { "count": 10, "median_ms": 23.295121267437935, "p95_ms": 24.609151761978865, "range_ms": [ 23.227659054100513, 24.609151761978865 ] }, "warm_median_decisions_per_second": 43.773012241422464 }, "8": { "warm": { "count": 10, "median_ms": 27.314365142956376, "p95_ms": 28.961639385670424, "range_ms": [ 27.18529710546136, 28.961639385670424 ] }, "cold": { "count": 10, "median_ms": 27.537896996363997, "p95_ms": 29.009283054620028, "range_ms": [ 27.268899139016867, 29.009283054620028 ] }, "warm_median_decisions_per_second": 292.88617758934004 }, "16": { "warm": { "count": 10, "median_ms": 29.842145973816514, "p95_ms": 30.099600087851286, "range_ms": [ 29.773924965411425, 30.099600087851286 ] }, "cold": { "count": 10, "median_ms": 29.994444688782096, "p95_ms": 32.27252420037985, "range_ms": [ 29.866321943700314, 32.27252420037985 ] }, "warm_median_decisions_per_second": 536.1544713988865 }, "32": { "warm": { "count": 10, "median_ms": 35.80718184821308, "p95_ms": 39.09342689439654, "range_ms": [ 35.35476792603731, 39.09342689439654 ] }, "cold": { "count": 10, "median_ms": 35.529671935364604, "p95_ms": 38.46615785732865, "range_ms": [ 35.424796398729086, 38.46615785732865 ] }, "warm_median_decisions_per_second": 893.6754681127449 }, "64": { "warm": { "count": 10, "median_ms": 57.56832659244537, "p95_ms": 59.25153801217675, "range_ms": [ 57.05864494666457, 59.25153801217675 ] }, "cold": { "count": 10, "median_ms": 57.97937046736479, "p95_ms": 59.614480938762426, "range_ms": [ 57.77294002473354, 59.614480938762426 ] }, "warm_median_decisions_per_second": 1111.7224312092242 } }, "performance": { "1": { "warm": { "count": 10, "median_ms": 24.860800243914127, "p95_ms": 26.331953704357147, "range_ms": [ 24.80240212753415, 26.331953704357147 ] }, "cold": { "count": 10, "median_ms": 24.9115196056664, "p95_ms": 25.805712677538395, "range_ms": [ 24.311278015375137, 25.805712677538395 ] }, "warm_median_decisions_per_second": 40.223966653881064 }, "8": { "warm": { "count": 10, "median_ms": 67.04071001149714, "p95_ms": 68.26437124982476, "range_ms": [ 66.94795889779925, 68.26437124982476 ] }, "cold": { "count": 10, "median_ms": 67.82145774923265, "p95_ms": 69.17453603819013, "range_ms": [ 67.67716398462653, 69.17453603819013 ] }, "warm_median_decisions_per_second": 119.3304784306139 }, "16": { "warm": { "count": 10, "median_ms": 142.1585502102971, "p95_ms": 143.4747572056949, "range_ms": [ 141.72939117997885, 143.4747572056949 ] }, "cold": { "count": 10, "median_ms": 143.65023071877658, "p95_ms": 145.14261297881603, "range_ms": [ 143.00691802054644, 145.14261297881603 ] }, "warm_median_decisions_per_second": 112.55038811475623 }, "32": { "warm": { "count": 10, "median_ms": 314.55507851205766, "p95_ms": 318.1276097893715, "range_ms": [ 313.771678134799, 318.1276097893715 ] }, "cold": { "count": 10, "median_ms": 317.7179682534188, "p95_ms": 319.81720635667443, "range_ms": [ 316.54817145317793, 319.81720635667443 ] }, "warm_median_decisions_per_second": 101.73099144153021 }, "64": { "warm": { "count": 10, "median_ms": 647.3779694642872, "p95_ms": 649.9390099197626, "range_ms": [ 644.2375518381596, 649.9390099197626 ] }, "cold": { "count": 10, "median_ms": 655.2308923564851, "p95_ms": 657.3728658258915, "range_ms": [ 652.8241420164704, 657.3728658258915 ] }, "warm_median_decisions_per_second": 98.86033046963391 } } }, "artifact": {}, "serving_configuration": null, "source_pins": { "kev_git": "bd058057ad0aa9df3dd6d14e3542c95a5ce367b2", "laya_git": "42626c348753fbb17572a813127df2278a1ec527", "jevbench_git": "c6004e008ffba24aec091261ca1a5c02f7324702", "kev_4b": "f780095d6511e6304868ff46ca2b74ee3eba29d8", "kev_9b": "6281032426a9ca3a08374d137bf9dfd8afd78e8a", "laya": "1c5edc17a7acd8701df6fc341c0d179f1c62c982", "laya_typed": "f9ab0b228f0fc0f14d873dbc99038f135c2da1b2", "typed_decisions": "ea9306458d6e9563628369a3d1e72e362fb381d2" }, "dataset_manifest_sha256": { "kev_v9": "3c4f0be94509a3612678bfd3a30fd99a8d0ca3c47ddfe7318075d95b2fa365e4", "jevbench": "0983ad3e73a34e7c55b02b5d09ed983aeccabef0253bb31be43a2e6230d58f7b" }, "inference_build": {} }, { "model": "Laya English", "precision": "float32 weights; bfloat16 autocast", "quality_percent": { "jevbench": 58.44155844155844, "kev_v9": 52.96367112810707, "kev_additions": 32.05128205128205, "typed": 36.1 }, "counts": { "jevbench": 231, "kev_v9": 1046, "kev_additions": 390, "typed": 2000 }, "calibration": { "jev_brier": 0.53208797, "jev_ece": 0.09038658008658014, "kev_brier": 0.5825189168117686, "kev_ece": 0.13014756076373651, "typed_teacher_brier_non_score": 0.32897452841925434, "typed_ece": 0.17523136709240314 }, "memory_gib": { "gpu": "NVIDIA RTX PRO 5000 Blackwell", "device_baseline_mib": 2.0, "load_vram_gib": 1.841796875, "idle_vram_gib": 1.919921875, "inference_vram_gib": 4.00390625, "load_ram_pss_gib": 3.2475242614746094, "idle_ram_pss_gib": 1.0939559936523438, "inference_ram_pss_gib": 2.0521316528320312 }, "performance": { "performance-short": { "1": { "warm": { "count": 10, "median_ms": 23.84014753624797, "p95_ms": 24.90654308348894, "range_ms": [ 23.538042791187763, 24.90654308348894 ] }, "cold": { "count": 10, "median_ms": 24.082106538116932, "p95_ms": 24.846371728926897, "range_ms": [ 24.05818784609437, 24.846371728926897 ] }, "warm_median_decisions_per_second": 41.946049137470354 }, "8": { "warm": { "count": 10, "median_ms": 28.08005828410387, "p95_ms": 30.12946993112564, "range_ms": [ 28.003456071019173, 30.12946993112564 ] }, "cold": { "count": 10, "median_ms": 28.088030638173223, "p95_ms": 29.704309068620205, "range_ms": [ 28.03805796429515, 29.704309068620205 ] }, "warm_median_decisions_per_second": 284.8996935497389 }, "16": { "warm": { "count": 10, "median_ms": 30.665626749396324, "p95_ms": 33.649255987256765, "range_ms": [ 30.55564919486642, 33.649255987256765 ] }, "cold": { "count": 10, "median_ms": 30.93404253013432, "p95_ms": 32.32662100344896, "range_ms": [ 30.611341819167137, 32.32662100344896 ] }, "warm_median_decisions_per_second": 521.7568233890727 }, "32": { "warm": { "count": 10, "median_ms": 36.21700336225331, "p95_ms": 38.06876204907894, "range_ms": [ 36.08020907267928, 38.06876204907894 ] }, "cold": { "count": 10, "median_ms": 36.47041297517717, "p95_ms": 38.297314662486315, "range_ms": [ 36.158771719783545, 38.297314662486315 ] }, "warm_median_decisions_per_second": 883.5628856403833 }, "64": { "warm": { "count": 10, "median_ms": 57.41862556897104, "p95_ms": 58.74502472579479, "range_ms": [ 57.02610267326236, 58.74502472579479 ] }, "cold": { "count": 10, "median_ms": 57.85714997909963, "p95_ms": 58.998359367251396, "range_ms": [ 57.67525406554341, 58.998359367251396 ] }, "warm_median_decisions_per_second": 1114.6208981112486 } }, "performance": { "1": { "warm": { "count": 10, "median_ms": 25.882226647809148, "p95_ms": 26.91890113055706, "range_ms": [ 25.42299311608076, 26.91890113055706 ] }, "cold": { "count": 10, "median_ms": 25.474075926467776, "p95_ms": 25.56223887950182, "range_ms": [ 25.444936007261276, 25.56223887950182 ] }, "warm_median_decisions_per_second": 38.63655216405606 }, "8": { "warm": { "count": 10, "median_ms": 45.27222109027207, "p95_ms": 47.11010167375207, "range_ms": [ 45.21714290603995, 47.11010167375207 ] }, "cold": { "count": 10, "median_ms": 45.3107540961355, "p95_ms": 46.22122226282954, "range_ms": [ 45.23011017590761, 46.22122226282954 ] }, "warm_median_decisions_per_second": 176.7088030438827 }, "16": { "warm": { "count": 10, "median_ms": 89.15169816464186, "p95_ms": 90.1275728829205, "range_ms": [ 88.40252412483096, 90.1275728829205 ] }, "cold": { "count": 10, "median_ms": 88.64646148867905, "p95_ms": 90.57755582034588, "range_ms": [ 88.46736513078213, 90.57755582034588 ] }, "warm_median_decisions_per_second": 179.4693800498542 }, "32": { "warm": { "count": 10, "median_ms": 191.60822709091008, "p95_ms": 194.2352126352489, "range_ms": [ 191.27637008205056, 194.2352126352489 ] }, "cold": { "count": 10, "median_ms": 191.81448104791343, "p95_ms": 193.55412013828754, "range_ms": [ 191.48092670366168, 193.55412013828754 ] }, "warm_median_decisions_per_second": 167.00744266485665 }, "64": { "warm": { "count": 10, "median_ms": 412.5161219853908, "p95_ms": 416.34345101192594, "range_ms": [ 411.5443374030292, 416.34345101192594 ] }, "cold": { "count": 10, "median_ms": 415.1305085979402, "p95_ms": 416.15893226116896, "range_ms": [ 411.9292129762471, 416.15893226116896 ] }, "warm_median_decisions_per_second": 155.1454515085996 } } }, "artifact": {}, "serving_configuration": null, "source_pins": { "kev_git": "bd058057ad0aa9df3dd6d14e3542c95a5ce367b2", "laya_git": "42626c348753fbb17572a813127df2278a1ec527", "jevbench_git": "c6004e008ffba24aec091261ca1a5c02f7324702", "kev_4b": "f780095d6511e6304868ff46ca2b74ee3eba29d8", "kev_9b": "6281032426a9ca3a08374d137bf9dfd8afd78e8a", "laya": "1c5edc17a7acd8701df6fc341c0d179f1c62c982", "laya_typed": "f9ab0b228f0fc0f14d873dbc99038f135c2da1b2", "typed_decisions": "ea9306458d6e9563628369a3d1e72e362fb381d2" }, "dataset_manifest_sha256": { "kev_v9": "3c4f0be94509a3612678bfd3a30fd99a8d0ca3c47ddfe7318075d95b2fa365e4", "jevbench": "0983ad3e73a34e7c55b02b5d09ed983aeccabef0253bb31be43a2e6230d58f7b" }, "inference_build": {} } ], "hardware_5070ti": { "gpu": "RTX 5070 Ti 16 GB", "prefix_positions": 65022, "image_positions": 1024, "decision_branches": 4, "chat_slots": 1, "kv": "Q8_0", "effective_memory_policy": "exclusive", "decision_cold_s": 25.001564278034493, "decision_warm_median_ms": 143.02524807862937, "decision_warm_samples": 3, "retrieval_passed": true, "short_generation_tps_median": 55.46540467494125, "short_generation_samples": 3, "short_output_tokens_per_sample": 512, "long_prompt_tokens": 62435, "long_output_tokens": 512, "long_generation_tps": 46.93347166614375, "long_prefill_tps": 2893.8818373839104, "long_ttft_s": 21.747452786192298, "long_total_s": 32.63961890107021, "smoke_peak_device_vram_gib": 15.0087890625, "smoke_peak_ram_pss_gib_including_startup": 12.247476577758789, "smoke_device_baseline_mib": 450.0 }, "jev": { "label": "Jev 1.13 (OpenRouter)", "model_revisions": [ "typesafe/jev-1.13-20260917" ], "quality": { "jevbench": 85.71428571428571, "kev_v9": 86.9980879541109, "kev_additions": 85.64102564102564, "typed": 73.8 }, "counts": { "jevbench": 231, "jevbench_correct": 198, "kev_v9": 1046, "kev_additions": 390, "typed_cases": 400, "typed": 2000 }, "calibration": { "jev_brier": 0.17946493506493508, "jev_ece": 0.04740259740259756, "kev_brier": 0.19142599481108083, "kev_ece": 0.030807501400235617, "typed_teacher_brier_non_score": 0.09788350992788927, "typed_ece": 0.03695434343434338 }, "coverage": { "requests": 1895, "jevbench": 231, "kev_v9": 1264, "typed_cases": 400, "typed_decisions": 2000, "unique_api_requests": 1840, "failed_requests": 0, "jevbench_schema_validity": 1.0, "contract_warnings": 0 }, "cost_usd": 0.04605409200000002, "input_tokens": 1096526, "output_tokens": 114470, "pins": { "kev_git": "bd058057ad0aa9df3dd6d14e3542c95a5ce367b2", "laya_git": "42626c348753fbb17572a813127df2278a1ec527", "jevbench_git": "c6004e008ffba24aec091261ca1a5c02f7324702", "kev_4b": "f780095d6511e6304868ff46ca2b74ee3eba29d8", "kev_9b": "6281032426a9ca3a08374d137bf9dfd8afd78e8a", "laya": "1c5edc17a7acd8701df6fc341c0d179f1c62c982", "laya_typed": "f9ab0b228f0fc0f14d873dbc99038f135c2da1b2", "typed_decisions": "ea9306458d6e9563628369a3d1e72e362fb381d2" }, "collection_date": "2026-09-21", "latency_scope": "OpenRouter round trip; not comparable to pod-local latency", "gpu_memory_gib": null, "system_memory_gib": null, "metric_scope": "JevBench public-subset accuracy; Kev-v9 knowable clean accuracy; typed teacher agreement.", "deduplication": "Identical payloads reuse one frozen response; all evaluation records remain scored." }, "fine_tuning_overlap_audit": { "by_suite": { "jevbench": { "evaluated_records": 231, "exact_normalized_input": 0, "exact_normalized_state": 0, "high_containment_state_candidates": 0 }, "kev-v9": { "evaluated_records": 1264, "exact_normalized_input": 0, "exact_normalized_state": 0, "high_containment_state_candidates": 0 }, "typed-decisions": { "evaluated_records": 400, "exact_normalized_input": 0, "exact_normalized_state": 0, "high_containment_state_candidates": 0 } }, "method": "Casefolded alphanumeric exact state/input matching; question IDs and label fields excluded; typed cases split into questions; 5-word state containment >=85% with minimum20 shingles.", "scope": "Checksum-verified private fine-tuning inputs. Does not audit Gemma pretraining or exclude conceptual/source-family overlap or development exposure. The training data and training pipeline remain private and are not distributed." } }