{ "schemaVersion": 1, "generatedUtc": "2026-08-26T20:08:16.158361+00:00", "reference": { "fileName": "Qwen3.8-27B-Quark-AWQ-MXFP4-native.gguf", "sha256": "1a37c48570811215fa1a9d2a493211293ca375d17287771e85d28642946e88fc", "standalonePpl": 5.8033 }, "method": { "domain": "general", "dataset": "Salesforce/wikitext", "tokenTarget": 32768, "corpusBytes": 131514, "corpusSha256": "5d38d98dce15f54e9a1a926187b6058e65cd8b3dd9b5cc2729b0a5fd249228b0", "referenceLogitsSha256": "604db2df7253cba18a6c0a569ed819ee3065d4ac20cd94ae72f4329e7bba8516", "commandShape": "llama-perplexity -m -ngl 0 -t 4 -c 2048 --file --kl-divergence-base --kl-divergence", "gpuUsed": false }, "results": [ { "name": "MQ-IQ4_XS_1", "fileName": "Qwen3.8-27B-Quark-MXFP4-MQ-IQ4_XS_1-Generic.gguf", "imatrix": "Generic", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 5.800778, "basePpl": 5.801511, "pplDifference": -0.000733, "kld": { "mean": 0.00094, "standardError": 2.9e-05, "maximum": 0.254027, "p99_9": 0.026253, "p99": 0.006447, "p95": 0.002801, "p90": 0.001907, "median": 0.000516 }, "tokenProbability": { "meanDeltaPercent": -0.013, "rmsDeltaPercent": 0.907, "sameTopPercent": 98.527 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "published" }, { "name": "UD-Q4_K_S", "fileName": "Qwen3.8-27B-Quark-MXFP4-UD-Q4_K_S-Unsloth.gguf", "imatrix": "Unsloth", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 5.80558, "basePpl": 5.801511, "pplDifference": 0.004069, "kld": { "mean": 0.00345, "standardError": 9.9e-05, "maximum": 0.602664, "p99_9": 0.191006, "p99": 0.034935, "p95": 0.011178, "p90": 0.006836, "median": 0.001342 }, "tokenProbability": { "meanDeltaPercent": -0.036, "rmsDeltaPercent": 1.809, "sameTopPercent": 97.471 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "published" }, { "name": "UD-IQ4_XS", "fileName": "Qwen3.8-27B-Quark-MXFP4-UD-IQ4_XS-Unsloth.gguf", "imatrix": "Unsloth", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 5.82498, "basePpl": 5.801511, "pplDifference": 0.023469, "kld": { "mean": 0.009018, "standardError": 0.000191, "maximum": 0.904095, "p99_9": 0.31015, "p99": 0.085121, "p95": 0.030207, "p90": 0.019024, "median": 0.003997 }, "tokenProbability": { "meanDeltaPercent": -0.077, "rmsDeltaPercent": 2.827, "sameTopPercent": 95.758 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "published" }, { "name": "UD-Q3_K_XL", "fileName": "Qwen3.8-27B-Quark-MXFP4-UD-Q3_K_XL-Unsloth.gguf", "imatrix": "Unsloth", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 5.88221, "basePpl": 5.801511, "pplDifference": 0.080698, "kld": { "mean": 0.022344, "standardError": 0.000524, "maximum": 3.909688, "p99_9": 0.577549, "p99": 0.203187, "p95": 0.07903, "p90": 0.049604, "median": 0.009937 }, "tokenProbability": { "meanDeltaPercent": -0.377, "rmsDeltaPercent": 4.405, "sameTopPercent": 93.672 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "published" }, { "name": "UD-IQ3_S", "fileName": "Qwen3.8-27B-Quark-MXFP4-UD-IQ3_S-Unsloth.gguf", "imatrix": "Unsloth", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 5.951552, "basePpl": 5.801511, "pplDifference": 0.150041, "kld": { "mean": 0.042218, "standardError": 0.000861, "maximum": 5.45832, "p99_9": 1.217797, "p99": 0.400938, "p95": 0.151289, "p90": 0.092036, "median": 0.018492 }, "tokenProbability": { "meanDeltaPercent": -0.711, "rmsDeltaPercent": 6.061, "sameTopPercent": 91.398 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "published" }, { "name": "MQ-IQ2_M_1", "fileName": "Qwen3.8-27B-Quark-MXFP4-MQ-IQ2_M_1-Generic.gguf", "imatrix": "Generic", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 6.04002, "basePpl": 5.801511, "pplDifference": 0.238509, "kld": { "mean": 0.058879, "standardError": 0.001352, "maximum": 4.972449, "p99_9": 2.321059, "p99": 0.597265, "p95": 0.195393, "p90": 0.119437, "median": 0.025196 }, "tokenProbability": { "meanDeltaPercent": -1.207, "rmsDeltaPercent": 7.161, "sameTopPercent": 89.951 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "published" }, { "name": "UD-IQ3_XXS", "fileName": "Qwen3.8-27B-Quark-MXFP4-UD-IQ3_XXS-Unsloth.gguf", "imatrix": "Unsloth", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 6.122065, "basePpl": 5.801511, "pplDifference": 0.320554, "kld": { "mean": 0.072327, "standardError": 0.001382, "maximum": 4.86586, "p99_9": 2.386696, "p99": 0.761119, "p95": 0.249649, "p90": 0.155595, "median": 0.031431 }, "tokenProbability": { "meanDeltaPercent": -1.164, "rmsDeltaPercent": 7.64, "sameTopPercent": 88.667 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "published" }, { "name": "MQ-IQ2_M_2", "fileName": "Qwen3.8-27B-Quark-MXFP4-MQ-IQ2_M_2-Generic.gguf", "imatrix": "Generic", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 6.266562, "basePpl": 5.801511, "pplDifference": 0.465051, "kld": { "mean": 0.102424, "standardError": 0.001765, "maximum": 6.221836, "p99_9": 2.665715, "p99": 0.950551, "p95": 0.335705, "p90": 0.21974, "median": 0.052264 }, "tokenProbability": { "meanDeltaPercent": -1.851, "rmsDeltaPercent": 9.458, "sameTopPercent": 85.709 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "published" }, { "name": "UD-Q2_K_XL", "fileName": "Qwen3.8-27B-Quark-MXFP4-UD-Q2_K_XL-Unsloth.gguf", "imatrix": "Unsloth", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 6.280284, "basePpl": 5.801511, "pplDifference": 0.478773, "kld": { "mean": 0.111015, "standardError": 0.001835, "maximum": 5.839242, "p99_9": 2.934114, "p99": 0.993161, "p95": 0.397965, "p90": 0.247337, "median": 0.052314 }, "tokenProbability": { "meanDeltaPercent": -2.07, "rmsDeltaPercent": 9.66, "sameTopPercent": 86.087 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "published" }, { "name": "UD-IQ2_XXS", "fileName": "Qwen3.8-27B-Quark-MXFP4-UD-IQ2_XXS-Unsloth.gguf", "imatrix": "Unsloth", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 17.87418, "basePpl": 5.801511, "pplDifference": 12.072669, "kld": { "mean": 1.172122, "standardError": 0.026974, "maximum": 19.836115, "p99_9": 18.650188, "p99": 16.602785, "p95": 11.67594, "p90": 1.495165, "median": 0.20467 }, "tokenProbability": { "meanDeltaPercent": -10.964, "rmsDeltaPercent": 26.023, "sameTopPercent": 71.261 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "removed", "removalReasonCode": "FAILED_QUALITY_FLOOR_AND_STRICT_DOMINANCE", "removalReason": "Removed from distribution after verification: KLD 1.172122 and PPL 17.87418 versus native PPL 5.801511. It is also strictly dominated by MQ-IQ2_XXS_1, which is smaller (8.22 GB) and has much lower KLD (0.321797)." }, { "name": "MQ-IQ2_XXS_1", "fileName": "Qwen3.8-27B-Quark-MXFP4-MQ-IQ2_XXS_1-Unsloth.gguf", "imatrix": "Unsloth", "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "candidatePpl": 7.704137, "basePpl": 5.801511, "pplDifference": 1.902626, "kld": { "mean": 0.321797, "standardError": 0.004179, "maximum": 9.210259, "p99_9": 5.867157, "p99": 2.483742, "p95": 1.134889, "p90": 0.741691, "median": 0.177414 }, "tokenProbability": { "meanDeltaPercent": -6.684, "rmsDeltaPercent": 17.362, "sameTopPercent": 76.729 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" }, "publicationStatus": "published" } ] }