{ "schemaVersion": 1, "snapshotId": "sha256-0fcc9fe3b6f0b0eda114b09322a9aab863a391b7888800fc0591d5a532ce5552", "dataAsOf": "2026-08-28", "defaults": { "read": 0.75, "write": 0.25, "hours": 24, "electricity": 0.15, "life": 5, "apiInput": 0.14, "apiOutput": 0.28 }, "records": [ { "recordKey": "rtx3090:qwen38_27b", "rigId": "rtx3090", "hardware": "RTX 3090", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 24, "readingTps": 705, "writingTps": 55.9, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 10, "asOf": "2026-08-23", "seoEligible": true, "sourceUrls": [ "https://jonidimo.github.io/qwen38-3090-benchmark/benchmark.html", "https://huggingface.co/Zynerji/Qwen3.8-27B-PristinelyUncensored-HOMEUSER-16-24", "https://github.com/sudoingX/qwen38-mtp" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rtx3090ti:qwen38_27b", "rigId": "rtx3090ti", "hardware": "RTX 3090 Ti", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 24, "readingTps": 1458.3333333333335, "writingTps": 69.75, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 3, "asOf": "2026-08-18", "seoEligible": true, "sourceUrls": [ "https://huggingface.co/Zynerji/Qwen3.8-27B-PristinelyUncensored-HOMEUSER-16-24" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rx7900xtx:qwen38_27b", "rigId": "rx7900xtx", "hardware": "AMD Radeon RX 7900 XTX", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 24, "readingTps": 877.84, "writingTps": 43.9, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 3, "asOf": "2026-08-17", "seoEligible": true, "sourceUrls": [ "https://github.com/sudoingX/qwen38-mtp", "https://smeltcore.com/recipes/qwen3-8-27b-on-rx-7900-xtx-128k-context-vision-chat-on-rocm-with-llama-cpp/" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "dual5060:qwen38_27b", "rigId": "dual5060", "hardware": "2× RTX 5060 Ti", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 32, "readingTps": 723, "writingTps": 65.9, "readingStatus": "measured", "writingStatus": "measured", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 8, "asOf": "2026-08-24", "seoEligible": true, "sourceUrls": [ "https://github.com/sudoingX/qwen38-mtp" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rtx5060:qwen38_27b", "rigId": "rtx5060", "hardware": "RTX 5060 Ti", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 16, "readingTps": 779.5100222717149, "writingTps": 27.4, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 2, "asOf": "2026-08-18", "seoEligible": true, "sourceUrls": [ "https://huggingface.co/Zynerji/Qwen3.8-27B-PristinelyUncensored-HOMEUSER-16-24" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "dual5070:qwen38_27b", "rigId": "dual5070", "hardware": "2× RTX 5070 Ti", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 32, "readingTps": 3250, "writingTps": 113.25, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 12, "asOf": "2026-08-19", "seoEligible": true, "sourceUrls": [ "https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4/discussions/16", "https://huggingface.co/eemin/Qwen3.8-27B-NVFP4-MTP-2x16GB" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rtx5070:qwen38_27b", "rigId": "rtx5070", "hardware": "RTX 5070 Ti", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 16, "readingTps": 1555.5555555555557, "writingTps": 50.8, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 2, "asOf": "2026-08-18", "seoEligible": true, "sourceUrls": [ "https://huggingface.co/Zynerji/Qwen3.8-27B-PristinelyUncensored-HOMEUSER-16-24" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "dual4080:qwen38_27b", "rigId": "dual4080", "hardware": "2× RTX 4080", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 32, "readingTps": 1700, "writingTps": 58, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": null, "seoEligible": false, "sourceUrls": [], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rtx4080super:qwen38_27b", "rigId": "rtx4080super", "hardware": "RTX 4080 Super", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 16, "readingTps": 1750, "writingTps": 43.6, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 2, "asOf": "2026-08-18", "seoEligible": true, "sourceUrls": [ "https://huggingface.co/Zynerji/Qwen3.8-27B-PristinelyUncensored-HOMEUSER-16-24" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rtx4090:qwen38_27b", "rigId": "rtx4090", "hardware": "RTX 4090", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 24, "readingTps": 2692.3076923076924, "writingTps": 68.8, "readingStatus": "measured", "writingStatus": "measured", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 6, "asOf": "2026-08-24", "seoEligible": true, "sourceUrls": [ "https://huggingface.co/Zynerji/Qwen3.8-27B-PristinelyUncensored-HOMEUSER-16-24", "https://github.com/sudoingX/qwen38-mtp" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rtx5090:qwen38_27b", "rigId": "rtx5090", "hardware": "RTX 5090", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 32, "readingTps": 3571.4285714285716, "writingTps": 132.5, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 13, "asOf": "2026-08-24", "seoEligible": true, "sourceUrls": [ "https://github.com/sudoingX/qwen38-mtp", "https://huggingface.co/Zynerji/Qwen3.8-27B-PristinelyUncensored-HOMEUSER-16-24" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "arcpro_b70:qwen38_27b", "rigId": "arcpro_b70", "hardware": "Intel Arc Pro B70", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 32, "readingTps": 424.8, "writingTps": 49.349999999999994, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 5, "asOf": "2026-08-25", "seoEligible": true, "sourceUrls": [ "https://huggingface.co/SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16", "https://jonathanmann.tech/blog/qwen38-intel-arc-b70/", "https://github.com/vllm-project/vllm-xpu-kernels/issues/510", "https://github.com/SergiioB/intel-arc-pro-b70-inference-cookbook/issues/6" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rtxpro5000:qwen38_27b", "rigId": "rtxpro5000", "hardware": "RTX PRO 5000 Blackwell", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 48, "readingTps": 6288, "writingTps": 112.5, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 1, "asOf": "2026-08-23", "seoEligible": true, "sourceUrls": [ "https://github.com/vllm-project/vllm/issues/53428" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rtxpro6000:qwen38_27b", "rigId": "rtxpro6000", "hardware": "RTX PRO 6000 Blackwell", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 96, "readingTps": 7756, "writingTps": 109.25, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 4, "asOf": "2026-08-25", "seoEligible": true, "sourceUrls": [ "https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000/", "https://github.com/sgl-project/sglang/issues/36118", "https://huggingface.co/lribeiro/Qwen3.8-27B-nvfp4-v6", "https://huggingface.co/pearsonkyle/Qwen3.8-27B-GPTQ-W4A16", "https://huggingface.co/Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored-GPTQ-MTP" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rtx_a4000:qwen38_27b", "rigId": "rtx_a4000", "hardware": "RTX A4000", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 16, "readingTps": 925.925925925926, "writingTps": 28.3, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 2, "asOf": "2026-08-18", "seoEligible": true, "sourceUrls": [ "https://huggingface.co/Zynerji/Qwen3.8-27B-PristinelyUncensored-HOMEUSER-16-24" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "m3ultra:qwen38_27b", "rigId": "m3ultra", "hardware": "Mac Studio M3 Ultra 512 GB", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 512, "readingTps": 462.5, "writingTps": 44.25, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 2, "asOf": "2026-08-23", "seoEligible": true, "sourceUrls": [ "https://omlx.ai/benchmarks/performance/5x3mzt9b", "https://github.com/sudoingX/qwen38-mtp" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "m4max:qwen38_27b", "rigId": "m4max", "hardware": "Mac Studio M4 Max 64 GB", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 64, "readingTps": 274, "writingTps": 53.3, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": null, "seoEligible": false, "sourceUrls": [], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "m4max128:qwen38_27b", "rigId": "m4max128", "hardware": "Mac Studio M4 Max 128 GB", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 128, "readingTps": null, "writingTps": 56.75, "readingStatus": "pending", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 7, "asOf": "2026-08-21", "seoEligible": true, "sourceUrls": [ "https://github.com/Weschera/Qwen3.8-27B-oMLX-MTP-Mac" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "m5max:qwen38_27b", "rigId": "m5max", "hardware": "MacBook Pro M5 Max 128 GB", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 128, "readingTps": 745, "writingTps": 59.2, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 1, "asOf": "2026-08-15", "seoEligible": true, "sourceUrls": [ "https://github.com/bluehawana/ollama-deepseek-v4-flash-iq2_m-mac-m5max" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "macmini_m6_32:qwen38_27b", "rigId": "macmini_m6_32", "hardware": "Mac mini M6 32 GB", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 32, "readingTps": 206, "writingTps": 16.4, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-15", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-mini/m6-chip-12-core-cpu-12-core-gpu-32gb-memory-1tb-storage" ], "estimateSource": "Projection from reported M5 Max Qwen3.8 rates, scaled to 170GB/s unified-memory bandwidth; not a Mac mini measurement", "estimateAnchors": [ { "rig": "m5max", "factor": 0.2768729641693811, "sourceBandwidth": 614, "targetBandwidth": 170 } ], "projectionMethod": null }, { "recordKey": "macmini_m5pro_64:qwen38_27b", "rigId": "macmini_m5pro_64", "hardware": "Mac mini M5 Pro 64 GB", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 64, "readingTps": 373, "writingTps": 29.6, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-15", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-mini/m5-pro-chip-18-core-cpu-20-core-gpu-64gb-memory-512gb-storage" ], "estimateSource": "Projection from reported M5 Max Qwen3.8 rates, scaled to 307GB/s unified-memory bandwidth; not a Mac mini measurement", "estimateAnchors": [ { "rig": "m5max", "factor": 0.5, "sourceBandwidth": 614, "targetBandwidth": 307 } ], "projectionMethod": null }, { "recordKey": "macstudio_m5max_64:qwen38_27b", "rigId": "macstudio_m5max_64", "hardware": "Mac Studio M5 Max 64 GB", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 64, "readingTps": 745, "writingTps": 59.2, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-15", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-studio/m5-max-chip-18-core-cpu-40-core-gpu-64gb-memory-1tb-storage" ], "estimateSource": "Projection anchored to the reported M5 Max Qwen3.8 source median; no Mac Studio measurement yet", "estimateAnchors": [ { "rig": "m5max", "factor": 1, "sourceBandwidth": 614, "targetBandwidth": 614 } ], "projectionMethod": null }, { "recordKey": "macstudio_m5max_128:qwen38_27b", "rigId": "macstudio_m5max_128", "hardware": "Mac Studio M5 Max 128 GB", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 128, "readingTps": 745, "writingTps": 59.2, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-15", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-studio/m5-max-chip-18-core-cpu-40-core-gpu-128gb-memory-1tb-storage" ], "estimateSource": "Projection anchored to the reported M5 Max Qwen3.8 source median; no Mac Studio measurement yet", "estimateAnchors": [ { "rig": "m5max", "factor": 1, "sourceBandwidth": 614, "targetBandwidth": 614 } ], "projectionMethod": null }, { "recordKey": "macstudio_m5ultra_96:qwen38_27b", "rigId": "macstudio_m5ultra_96", "hardware": "Mac Studio M5 Ultra 96 GB", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 96, "readingTps": 1075, "writingTps": 91, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-23", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-studio/m5-ultra-chip-30-core-cpu-64-core-gpu-96gb-memory-1tb-storage" ], "estimateSource": "Conservative midpoint of bandwidth-scaled M3 Ultra and M5 Max Qwen3.8 source medians; not a Mac Studio M5 Ultra measurement", "estimateAnchors": [ { "rig": "m5max", "factor": 1.9543973941368078, "sourceBandwidth": 614, "targetBandwidth": 1200 }, { "rig": "m3ultra", "factor": 1.5, "sourceBandwidth": 800, "targetBandwidth": 1200 } ], "projectionMethod": null }, { "recordKey": "macstudio_m5ultra_256:qwen38_27b", "rigId": "macstudio_m5ultra_256", "hardware": "Mac Studio M5 Ultra 256 GB", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 256, "readingTps": 1075, "writingTps": 91, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-23", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-studio/m5-ultra-chip-36-core-cpu-80-core-gpu-256gb-memory-1tb-storage" ], "estimateSource": "Conservative midpoint of bandwidth-scaled M3 Ultra and M5 Max Qwen3.8 source medians; not a Mac Studio M5 Ultra measurement", "estimateAnchors": [ { "rig": "m5max", "factor": 1.9543973941368078, "sourceBandwidth": 614, "targetBandwidth": 1200 }, { "rig": "m3ultra", "factor": 1.5, "sourceBandwidth": 800, "targetBandwidth": 1200 } ], "projectionMethod": null }, { "recordKey": "dgx:qwen38_27b", "rigId": "dgx", "hardware": "DGX Spark", "modelId": "qwen38_27b", "model": "Qwen3.8 27B", "variantId": "qwen38_27b_c1_personal", "memoryGB": 128, "readingTps": 1225.5, "writingTps": 44.0325, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 5, "asOf": "2026-08-24", "seoEligible": true, "sourceUrls": [ "https://github.com/hasso5703/dgx-spark-qwen38", "https://github.com/0xBakeer/Qwen3.8-27B-4-bit-on-a-single-DGX-Spark/blob/main/RESULTS.md" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "rtx3090:qwen38_flash_next", "rigId": "rtx3090", "hardware": "RTX 3090", "modelId": "qwen38_flash_next", "model": "Qwen3.8 Flash Next", "variantId": "qwen38_flash_next_c1_personal", "memoryGB": 24, "readingTps": 116, "writingTps": 5.791, "readingStatus": "estimate", "writingStatus": "reported", "readingEstimateLow": 100, "readingEstimateHigh": 150, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": true, "partialEstimate": true, "sourceBacked": true, "sourceObservationCount": 1, "asOf": "2026-08-28", "seoEligible": true, "sourceUrls": [ "https://github.com/FlashML-org/FreeToken/pull/232", "https://github.com/FlashML-org/FreeToken/pull/257#issuecomment-5458419210" ], "estimateSource": "RTX 3090 estimated warm 6.5K effective prefill: A100 1,367.25 pp × target FreeToken decode 5.791 / 68.30 = 116 pp; expected range 100–150 pp. Target prefill is not measured.", "estimateAnchors": [ { "rig": "a100_sxm4_80gb_freetoken", "anchorName": "NVIDIA A100 SXM4 80GB", "factor": 0.08478770131771597, "sourcePrefill": 1367.25, "sourceDecode": 68.3, "targetDecode": 5.791, "targetLow": 100, "targetHigh": 150, "sourcePromptTokens": 6500, "sourceUrl": "https://github.com/FlashML-org/FreeToken/pull/257#issuecomment-5458419210", "sourceDate": "28 Aug 2026", "sourceCondition": "FreeToken official FP8 · warm 6.5K prompt · zero cached prompt tokens; 1,367.25 effective pp and 68.30 mean decode are different source aggregates.", "formula": "1367.25 × 5.791 / 68.30 = 116 pp" } ], "projectionMethod": "freetoken-decode-normalized-prefill" }, { "recordKey": "rtx4090:qwen38_flash_next", "rigId": "rtx4090", "hardware": "RTX 4090", "modelId": "qwen38_flash_next", "model": "Qwen3.8 Flash Next", "variantId": "qwen38_flash_next_c1_personal", "memoryGB": 24, "readingTps": 721, "writingTps": 36, "readingStatus": "estimate", "writingStatus": "reported", "readingEstimateLow": 650, "readingEstimateHigh": 850, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": true, "partialEstimate": true, "sourceBacked": true, "sourceObservationCount": 1, "asOf": "2026-08-28", "seoEligible": true, "sourceUrls": [ "https://github.com/FlashML-org/FreeToken/pull/257", "https://github.com/FlashML-org/FreeToken/pull/257#issuecomment-5458419210" ], "estimateSource": "RTX 4090 estimated warm 6.5K effective prefill: A100 1,367.25 pp × target FreeToken decode 36 / 68.30 = 721 pp; expected range 650–850 pp. Target prefill is not measured.", "estimateAnchors": [ { "rig": "a100_sxm4_80gb_freetoken", "anchorName": "NVIDIA A100 SXM4 80GB", "factor": 0.527086383601757, "sourcePrefill": 1367.25, "sourceDecode": 68.3, "targetDecode": 36, "targetLow": 650, "targetHigh": 850, "sourcePromptTokens": 6500, "sourceUrl": "https://github.com/FlashML-org/FreeToken/pull/257#issuecomment-5458419210", "sourceDate": "28 Aug 2026", "sourceCondition": "FreeToken official FP8 · warm 6.5K prompt · zero cached prompt tokens; 1,367.25 effective pp and 68.30 mean decode are different source aggregates.", "formula": "1367.25 × 36 / 68.30 = 721 pp" } ], "projectionMethod": "freetoken-decode-normalized-prefill" }, { "recordKey": "rtx5090:qwen38_flash_next", "rigId": "rtx5090", "hardware": "RTX 5090", "modelId": "qwen38_flash_next", "model": "Qwen3.8 Flash Next", "variantId": "qwen38_flash_next_c1_personal", "memoryGB": 32, "readingTps": 1301, "writingTps": 65, "readingStatus": "estimate", "writingStatus": "reported", "readingEstimateLow": 1200, "readingEstimateHigh": 1400, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": true, "partialEstimate": true, "sourceBacked": true, "sourceObservationCount": 1, "asOf": "2026-08-28", "seoEligible": true, "sourceUrls": [ "https://github.com/FlashML-org/FreeToken/pull/257", "https://github.com/FlashML-org/FreeToken/pull/257#issuecomment-5458419210" ], "estimateSource": "RTX 5090 estimated warm 6.5K effective prefill: A100 1,367.25 pp × target FreeToken decode 65 / 68.30 = 1301 pp; expected range 1,200–1,400 pp. Target prefill is not measured.", "estimateAnchors": [ { "rig": "a100_sxm4_80gb_freetoken", "anchorName": "NVIDIA A100 SXM4 80GB", "factor": 0.951683748169839, "sourcePrefill": 1367.25, "sourceDecode": 68.3, "targetDecode": 65, "targetLow": 1200, "targetHigh": 1400, "sourcePromptTokens": 6500, "sourceUrl": "https://github.com/FlashML-org/FreeToken/pull/257#issuecomment-5458419210", "sourceDate": "28 Aug 2026", "sourceCondition": "FreeToken official FP8 · warm 6.5K prompt · zero cached prompt tokens; 1,367.25 effective pp and 68.30 mean decode are different source aggregates.", "formula": "1367.25 × 65 / 68.30 = 1301 pp" } ], "projectionMethod": "freetoken-decode-normalized-prefill" }, { "recordKey": "apple_m5max_macbook_64_atomic:qwen38_flash_next", "rigId": "apple_m5max_macbook_64_atomic", "hardware": "MacBook Pro M5 Max 64 GB", "modelId": "qwen38_flash_next", "model": "Qwen3.8 Flash Next", "variantId": "qwen38_flash_next_c1_personal", "memoryGB": 64, "readingTps": 517.9, "writingTps": 36, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 1, "asOf": "2026-08-28", "seoEligible": true, "sourceUrls": [ "https://huggingface.co/AtomicChat/Qwen3.8-Flash-Next-GGUF" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null }, { "recordKey": "macmini_m5pro_64:qwen38_flash_next", "rigId": "macmini_m5pro_64", "hardware": "Mac mini M5 Pro 64 GB", "modelId": "qwen38_flash_next", "model": "Qwen3.8 Flash Next", "variantId": "qwen38_flash_next_c1_personal", "memoryGB": 64, "readingTps": 259, "writingTps": 18, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-28", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-mini/m5-pro-chip-18-core-cpu-20-core-gpu-64gb-memory-512gb-storage" ], "estimateSource": "Mac mini M5 Pro Flash Next projection from the source-backed M5 Max 64GB result; capped by the slowest CPU-core, GPU-core, or unified-memory-bandwidth ratio. Not an exact-machine measurement.", "estimateAnchors": [ { "rig": "apple_m5max_macbook_64_atomic", "factor": 0.5, "sourceBandwidth": 614, "targetBandwidth": 307, "sourceCpuCores": 18, "targetCpuCores": 18, "sourceGpuCores": 40, "targetGpuCores": 20, "limiting": [ "gpu", "bandwidth" ] } ], "projectionMethod": "conservative-hardware-ratio" }, { "recordKey": "macstudio_m5max_64:qwen38_flash_next", "rigId": "macstudio_m5max_64", "hardware": "Mac Studio M5 Max 64 GB", "modelId": "qwen38_flash_next", "model": "Qwen3.8 Flash Next", "variantId": "qwen38_flash_next_c1_personal", "memoryGB": 64, "readingTps": 517.9, "writingTps": 36, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-28", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-studio/m5-max-chip-18-core-cpu-40-core-gpu-64gb-memory-1tb-storage" ], "estimateSource": "Mac Studio M5 Max Flash Next projection from the source-backed M5 Max 64GB result; capped by the slowest CPU-core, GPU-core, or unified-memory-bandwidth ratio. Not an exact-machine measurement.", "estimateAnchors": [ { "rig": "apple_m5max_macbook_64_atomic", "factor": 1, "sourceBandwidth": 614, "targetBandwidth": 614, "sourceCpuCores": 18, "targetCpuCores": 18, "sourceGpuCores": 40, "targetGpuCores": 40, "limiting": [ "cpu", "gpu", "bandwidth" ] } ], "projectionMethod": "conservative-hardware-ratio" }, { "recordKey": "macstudio_m5max_128:qwen38_flash_next", "rigId": "macstudio_m5max_128", "hardware": "Mac Studio M5 Max 128 GB", "modelId": "qwen38_flash_next", "model": "Qwen3.8 Flash Next", "variantId": "qwen38_flash_next_c1_personal", "memoryGB": 128, "readingTps": 517.9, "writingTps": 36, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-28", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-studio/m5-max-chip-18-core-cpu-40-core-gpu-128gb-memory-1tb-storage" ], "estimateSource": "Mac Studio M5 Max Flash Next projection from the source-backed M5 Max 64GB result; capped by the slowest CPU-core, GPU-core, or unified-memory-bandwidth ratio. Not an exact-machine measurement.", "estimateAnchors": [ { "rig": "apple_m5max_macbook_64_atomic", "factor": 1, "sourceBandwidth": 614, "targetBandwidth": 614, "sourceCpuCores": 18, "targetCpuCores": 18, "sourceGpuCores": 40, "targetGpuCores": 40, "limiting": [ "cpu", "gpu", "bandwidth" ] } ], "projectionMethod": "conservative-hardware-ratio" }, { "recordKey": "macstudio_m5ultra_96:qwen38_flash_next", "rigId": "macstudio_m5ultra_96", "hardware": "Mac Studio M5 Ultra 96 GB", "modelId": "qwen38_flash_next", "model": "Qwen3.8 Flash Next", "variantId": "qwen38_flash_next_c1_personal", "memoryGB": 96, "readingTps": 828.6, "writingTps": 57.6, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-28", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-studio/m5-ultra-chip-30-core-cpu-64-core-gpu-96gb-memory-1tb-storage" ], "estimateSource": "Mac Studio M5 Ultra 64-core Flash Next projection from the source-backed M5 Max 64GB result; capped by the slowest CPU-core, GPU-core, or unified-memory-bandwidth ratio. Not an exact-machine measurement.", "estimateAnchors": [ { "rig": "apple_m5max_macbook_64_atomic", "factor": 1.6, "sourceBandwidth": 614, "targetBandwidth": 1200, "sourceCpuCores": 18, "targetCpuCores": 30, "sourceGpuCores": 40, "targetGpuCores": 64, "limiting": [ "gpu" ] } ], "projectionMethod": "conservative-hardware-ratio" }, { "recordKey": "macstudio_m5ultra_256:qwen38_flash_next", "rigId": "macstudio_m5ultra_256", "hardware": "Mac Studio M5 Ultra 256 GB", "modelId": "qwen38_flash_next", "model": "Qwen3.8 Flash Next", "variantId": "qwen38_flash_next_c1_personal", "memoryGB": 256, "readingTps": 1012.2, "writingTps": 70.4, "readingStatus": "estimate", "writingStatus": "estimate", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Estimate", "estimate": true, "partialEstimate": false, "sourceBacked": false, "sourceObservationCount": 0, "asOf": "2026-08-28", "seoEligible": true, "sourceUrls": [ "https://www.apple.com/shop/buy-mac/mac-studio/m5-ultra-chip-36-core-cpu-80-core-gpu-256gb-memory-1tb-storage" ], "estimateSource": "Mac Studio M5 Ultra 80-core Flash Next projection from the source-backed M5 Max 64GB result; capped by the slowest CPU-core, GPU-core, or unified-memory-bandwidth ratio. Not an exact-machine measurement.", "estimateAnchors": [ { "rig": "apple_m5max_macbook_64_atomic", "factor": 1.9543973941368078, "sourceBandwidth": 614, "targetBandwidth": 1200, "sourceCpuCores": 18, "targetCpuCores": 36, "sourceGpuCores": 40, "targetGpuCores": 80, "limiting": [ "bandwidth" ] } ], "projectionMethod": "conservative-hardware-ratio" }, { "recordKey": "dgx:qwen38_flash_next", "rigId": "dgx", "hardware": "DGX Spark", "modelId": "qwen38_flash_next", "model": "Qwen3.8 Flash Next", "variantId": "qwen38_flash_next_c1_personal", "memoryGB": 128, "readingTps": 540.065, "writingTps": 21.1, "readingStatus": "reported", "writingStatus": "reported", "readingEstimateLow": null, "readingEstimateHigh": null, "writingEstimateLow": null, "writingEstimateHigh": null, "evidenceStatus": "Median", "estimate": false, "partialEstimate": false, "sourceBacked": true, "sourceObservationCount": 6, "asOf": "2026-08-28", "seoEligible": true, "sourceUrls": [ "https://github.com/0xBakeer/qwen38-flash-next-spark", "https://huggingface.co/Baekpica/Qwen3.8-Flash-Next-Mixed-Quant-SSD-PLE-GGUF" ], "estimateSource": null, "estimateAnchors": [], "projectionMethod": null } ] }