{ "schemaVersion": 1, "operationId": "20260825150522-16dbda5f2d", "checkedAt": "2026-08-25T15:24:13Z", "status": "pinned-source-zero-hardware-m5-ultra-capacity-and-runtime-readiness-audit", "sourceReceipts": [ { "id": "apple-mac-studio-launch", "url": "https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/", "status": 200, "bytes": 227495, "sha256": "4f1ac7406bb5caf77bc0bd2b45dda18798f18e8e6f8d1468e593337954431659" }, { "id": "apple-m5-ultra-launch", "url": "https://www.apple.com/newsroom/2026/08/apple-introduces-m6-and-m5-ultra-for-a-big-leap-in-performance-and-ai-compute/", "status": 200, "bytes": 187800, "sha256": "04ea7a567e0f9c654008d4cb31a4f46d6070e5dc59f2843e099bc12f5900a0ad" }, { "id": "zai-glm52-release", "url": "https://z.ai/blog/glm-5.2", "status": 200, "bytes": 598, "sha256": "63e94b9e11a2db64243ff18418011577f2a98ca7851ce9dcb0e97cfe34cc245a" }, { "id": "zhipu-research-index", "url": "https://www.zhipuai.cn/zh/research", "status": 200, "bytes": 1236461, "sha256": "188772a5cb6b65eb65f02024d89d153b692ef3d3cb2c98ddd233352c0dd6ac80" }, { "id": "glm52-config-pinned", "url": "https://huggingface.co/zai-org/GLM-5.2/resolve/b4734de4facf877f85769a911abafc5283eab3d9/config.json", "status": 200, "bytes": 3732, "sha256": "185f93ee6d12548e16a847e279dc0c3c90b1524c970b0866b42fb545747d859a" }, { "id": "unsloth-tree-pinned", "url": "https://huggingface.co/api/models/unsloth/GLM-5.2-GGUF/tree/abc55e72527792c6e77069c99b4cb7de16fa9f23?recursive=true&limit=1000", "status": 200, "bytes": 86059, "sha256": "18129a02853d33740970a78c8d4b77a595aa8699a3dee84959ef56dbec326932" }, { "id": "unsloth-model-card-pinned", "url": "https://huggingface.co/unsloth/GLM-5.2-GGUF/resolve/abc55e72527792c6e77069c99b4cb7de16fa9f23/README.md", "status": 200, "bytes": 8048, "sha256": "f00d59230da2c2b25ddab67e9254b164fa71928c67b05e84ffd89eb5891e7119" }, { "id": "lm-studio-glm52", "url": "https://lmstudio.ai/models/glm-5.2", "status": 200, "bytes": 98685, "sha256": "9a50bed7ad4502f145c700e43ad2abea7a30ef8943cac8da22180745fa781e3d" }, { "id": "llama-cpp-pr-25407", "url": "https://api.github.com/repos/ggml-org/llama.cpp/pulls/25407", "status": 200, "bytes": 21071, "sha256": "0ac50f10ac64f339af044643e33a86309cd40db7b297f5905490f56ecf9ce753" }, { "id": "llama-cpp-support-commit", "url": "https://api.github.com/repos/ggml-org/llama.cpp/commits/88bfee1429a2dfacec65d1b0c0852eb327991865", "status": 200, "bytes": 40852, "sha256": "bec4311b92596d4c8e0f56137bb0948e7948d0cbda90b3ddb04a4c6e33bdff25" }, { "id": "mlx-lm-issue-1418", "url": "https://api.github.com/repos/ml-explore/mlx-lm/issues/1418", "status": 200, "bytes": 4269, "sha256": "55cc1a323ee60ee9a6d877861c26cc14116ab120e608d7edbe39d4b060ae44bd" }, { "id": "mlx-lm-pr-1419", "url": "https://api.github.com/repos/ml-explore/mlx-lm/pulls/1419", "status": 200, "bytes": 18064, "sha256": "e41775b8d154c45f0c602353ba15d45594cd535908f265bfeb3efe8546ea3ee7" }, { "id": "mlx-lm-release-v0313", "url": "https://api.github.com/repos/ml-explore/mlx-lm/releases/tags/v0.31.3", "status": 200, "bytes": 4682, "sha256": "6edaaf0a4dfdfc8eaa1997830e569a748480a4efcf84b831c5d49c2e98df8303" } ], "revisions": { "glm52": "b4734de4facf877f85769a911abafc5283eab3d9", "unslothGguf": "abc55e72527792c6e77069c99b4cb7de16fa9f23", "llamaCppGlm52Support": "88bfee1429a2dfacec65d1b0c0852eb327991865", "mlxLmStable": "v0.31.3", "mlxLmOpenSupportPrHead": "90b5958aeb14eaf6c1400b81f4da8fcfc2834ec6" }, "hardware": { "m5MaxMaxUnifiedMemoryBytes": 128000000000, "m5MaxMemoryBandwidthBytesPerSecond": 614000000000, "m5UltraAdvertisedUnifiedMemoryBytes": 512000000000, "m5UltraMemoryBandwidthBytesPerSecond": 1200000000000, "m5UltraCpuCoresMax": 36, "m5UltraGpuCoresMax": 80, "m5UltraStartingPriceUsd": 5499, "generalAvailabilityDate": "2026-09-22", "memory512Availability": "late October 2026", "clusterSystems": 4, "clusterVendorSpeedupMax": 3, "capacityConvention": "Use Apple's literal 512 GB label as a conservative 512,000,000,000-byte planning ceiling; actual allocatable process memory was not available to measure." }, "model": { "artifact": "zai-org/GLM-5.2", "architecture": "GlmMoeDsaForCausalLM", "modelType": "glm_moe_dsa", "parameters": 753000000000, "activeParametersApprox": 40000000000, "decoderLayers": 78, "kvLoraRank": 512, "qkRopeHeadDim": 64, "maxPositionEmbeddings": 1048576, "indexTopK": 2048, "routedExperts": 256, "expertsPerToken": 8 }, "cacheRows": [ { "bytesPerCachedValue": 1, "bytesPerToken": 44928, "fullContextBytes": 47110422528, "fullContextGB": 47.1104, "fullContextGiB": 43.875 }, { "bytesPerCachedValue": 2, "bytesPerToken": 89856, "fullContextBytes": 94220845056, "fullContextGB": 94.2208, "fullContextGiB": 87.75 } ], "artifactRows": [ { "name": "BF16", "publisher": "Unsloth GGUF", "bytes": 1507988023008, "sizeGB": 1507.988, "sizeGiB": 1404.4233, "fitsM5Max128GBByFileBytes": false, "fitsM5Ultra512GBByFileBytes": false, "m5UltraGrossHeadroomBytes": -995988023008, "m5UltraGrossHeadroomGB": -995.988, "m5UltraGrossHeadroomGiB": -927.5861, "fullContextFitAtOneByteValuesBeforeOtherOverhead": false, "fullContextFitAtTwoByteValuesBeforeOtherOverhead": false, "theoreticalMaxTokensAtOneByteValuesBeforeOtherOverhead": 0, "theoreticalMaxTokensAtTwoByteValuesBeforeOtherOverhead": 0 }, { "name": "Q8_0", "publisher": "Unsloth GGUF", "bytes": 801357672256, "sizeGB": 801.3577, "sizeGiB": 746.3225, "fitsM5Max128GBByFileBytes": false, "fitsM5Ultra512GBByFileBytes": false, "m5UltraGrossHeadroomBytes": -289357672256, "m5UltraGrossHeadroomGB": -289.3577, "m5UltraGrossHeadroomGiB": -269.4853, "fullContextFitAtOneByteValuesBeforeOtherOverhead": false, "fullContextFitAtTwoByteValuesBeforeOtherOverhead": false, "theoreticalMaxTokensAtOneByteValuesBeforeOtherOverhead": 0, "theoreticalMaxTokensAtTwoByteValuesBeforeOtherOverhead": 0 }, { "name": "UD-Q5_K_M", "publisher": "Unsloth GGUF", "bytes": 560830479904, "sizeGB": 560.8305, "sizeGiB": 522.3141, "fitsM5Max128GBByFileBytes": false, "fitsM5Ultra512GBByFileBytes": false, "m5UltraGrossHeadroomBytes": -48830479904, "m5UltraGrossHeadroomGB": -48.8305, "m5UltraGrossHeadroomGiB": -45.4769, "fullContextFitAtOneByteValuesBeforeOtherOverhead": false, "fullContextFitAtTwoByteValuesBeforeOtherOverhead": false, "theoreticalMaxTokensAtOneByteValuesBeforeOtherOverhead": 0, "theoreticalMaxTokensAtTwoByteValuesBeforeOtherOverhead": 0 }, { "name": "UD-Q4_K_M", "publisher": "Unsloth GGUF", "bytes": 465825525088, "sizeGB": 465.8255, "sizeGiB": 433.8338, "fitsM5Max128GBByFileBytes": false, "fitsM5Ultra512GBByFileBytes": true, "m5UltraGrossHeadroomBytes": 46174474912, "m5UltraGrossHeadroomGB": 46.1745, "m5UltraGrossHeadroomGiB": 43.0033, "fullContextFitAtOneByteValuesBeforeOtherOverhead": false, "fullContextFitAtTwoByteValuesBeforeOtherOverhead": false, "theoreticalMaxTokensAtOneByteValuesBeforeOtherOverhead": 1027743, "theoreticalMaxTokensAtTwoByteValuesBeforeOtherOverhead": 513871 }, { "name": "UD-IQ4_XS", "publisher": "Unsloth GGUF", "bytes": 365313223776, "sizeGB": 365.3132, "sizeGiB": 340.2245, "fitsM5Max128GBByFileBytes": false, "fitsM5Ultra512GBByFileBytes": true, "m5UltraGrossHeadroomBytes": 146686776224, "m5UltraGrossHeadroomGB": 146.6868, "m5UltraGrossHeadroomGiB": 136.6127, "fullContextFitAtOneByteValuesBeforeOtherOverhead": true, "fullContextFitAtTwoByteValuesBeforeOtherOverhead": true, "theoreticalMaxTokensAtOneByteValuesBeforeOtherOverhead": 3264930, "theoreticalMaxTokensAtTwoByteValuesBeforeOtherOverhead": 1632465 }, { "name": "UD-Q3_K_M", "publisher": "Unsloth GGUF", "bytes": 342735510656, "sizeGB": 342.7355, "sizeGiB": 319.1973, "fitsM5Max128GBByFileBytes": false, "fitsM5Ultra512GBByFileBytes": true, "m5UltraGrossHeadroomBytes": 169264489344, "m5UltraGrossHeadroomGB": 169.2645, "m5UltraGrossHeadroomGiB": 157.6398, "fullContextFitAtOneByteValuesBeforeOtherOverhead": true, "fullContextFitAtTwoByteValuesBeforeOtherOverhead": true, "theoreticalMaxTokensAtOneByteValuesBeforeOtherOverhead": 3767461, "theoreticalMaxTokensAtTwoByteValuesBeforeOtherOverhead": 1883730 }, { "name": "UD-IQ2_M", "publisher": "Unsloth GGUF", "bytes": 238577580768, "sizeGB": 238.5776, "sizeGiB": 222.1927, "fitsM5Max128GBByFileBytes": false, "fitsM5Ultra512GBByFileBytes": true, "m5UltraGrossHeadroomBytes": 273422419232, "m5UltraGrossHeadroomGB": 273.4224, "m5UltraGrossHeadroomGiB": 254.6445, "fullContextFitAtOneByteValuesBeforeOtherOverhead": true, "fullContextFitAtTwoByteValuesBeforeOtherOverhead": true, "theoreticalMaxTokensAtOneByteValuesBeforeOtherOverhead": 6085791, "theoreticalMaxTokensAtTwoByteValuesBeforeOtherOverhead": 3042895 } ], "q4FullContextBoundary": { "q4PlusOneByteCacheBytes": 512935947616, "q4PlusOneByteCacheGB": 512.9359, "exceedsConservative512GBByGB": 0.9359, "interpretation": "Even the one-byte latent-cache estimate plus Q4 exceeds the conservative 512 GB ceiling before macOS, runtime, index data, or workspaces." }, "iq4FullContextBoundary": { "iq4PlusTwoByteCacheBytes": 459534068832, "iq4PlusTwoByteCacheGB": 459.5341, "grossRemainingGB": 52.4659, "interpretation": "IQ4 plus the two-byte latent-cache estimate fits the conservative ceiling on paper, but the remaining amount is not a measured runtime reserve or quality guarantee." }, "runtimeSupport": [ { "runtime": "llama.cpp", "checkedBuild": "source at and after 88bfee1429a2dfacec65d1b0c0852eb327991865", "architectureSupport": "merged", "m5UltraGlm52Benchmark": "not available", "decision": "eligible for a pinned canary after hardware ships; source support is not a speed, quality, or memory guarantee" }, { "runtime": "MLX-LM", "checkedBuild": "v0.31.3", "architectureSupport": "blocked for GLM-5.2 IndexShare", "m5UltraGlm52Benchmark": "not available", "decision": "reject the stable path; issue 1418 and PR 1419 were still open" }, { "runtime": "LM Studio", "checkedBuild": "public model catalog on 2026-08-25", "architectureSupport": "GLM-5.2 listed as Cloud/Bionic, not a local download", "m5UltraGlm52Benchmark": "Apple model unspecified", "decision": "do not transfer Apple's LM Studio prompt-processing multiplier to local GLM-5.2" }, { "runtime": "Apple Core AI / Core ML", "checkedBuild": "launch announcement", "architectureSupport": "no GLM-5.2 conversion or validation path published", "m5UltraGlm52Benchmark": "not available", "decision": "unverified for this model" } ], "appleClaimBoundary": { "lmStudioPromptProcessingMultiplierVsM3Ultra": 4, "testedModelNamed": false, "glm52Named": false, "clusterModelNamed": false, "decision": "Treat all multipliers as Apple-reported hardware/application evidence, not GLM-5.2 throughput." }, "runtimeBoundary": { "m5UltraHardwareAvailableToEditorialTeam": false, "m5UltraShipmentsStarted": false, "modelCalls": 0, "localGpuRuns": 0, "localMacRuns": 0, "downloadsOfModelWeights": 0, "claimsAllowed": "capacity arithmetic, source support state, and future canary design only" }, "aiHot": { "fingerprint": "f1-d891e4a77b7ff874", "batchPath": "tmp/aihot-seo-operations/batches/20260825150522-16dbda5f2d.json", "batchSha256": "ca6e9e9ca44689d981ac1cd0d2160e1af0e0ccf877bf456b961037b961266755", "items": [ { "id": "cmt8qee503kvnro73qp7543yv", "permalink": "https://aihot.virxact.com/items/cmt8qee503kvnro73qp7543yv", "classification": "publish", "reason": "The 512GB M5 Ultra announcement creates a direct, expensive GLM-5.2 capacity-versus-runtime decision that current sources do not answer safely." }, { "id": "cmt8qee503kvpro73kzthenuu", "permalink": "https://aihot.virxact.com/items/cmt8qee503kvpro73kzthenuu", "classification": "duplicate-event", "reason": "It is a second Apple announcement for the same M5 Ultra hardware event and does not justify another canonical." }, { "id": "cmt8bo2db30oaro735zu90tig", "permalink": "https://aihot.virxact.com/items/cmt8bo2db30oaro735zu90tig", "classification": "weak-glm-link", "reason": "FORGE is useful recommendation-security research, but the checked primary source did not establish a direct GLM-5.2 implementation or decision." } ] }, "serp": { "requestId": "serpapi-3068d10ff30b45ba885245d35cd0dd94", "query": "GLM-5.2 M5 Ultra Mac Studio", "locale": "United States / English", "requestedResults": 10, "organicResults": 7, "requestCost": 1, "value": "high-value", "evidencePath": "docs/evidence/glm-5-2-m5-ultra-2026-08-25/serp-glm-5-2-m5-ultra.json", "sha256": "4ea6423f3a915c265bf93d165590784cc88df40e3f590603ec62f10d74e042f2", "decisionImpact": "The SERP showed a distinct Mac purchase-decision intent, conflicting memory/runtime claims, and no result about the newly announced M5 Ultra." }, "overlapAudit": { "prepublicationCanonicalCount": 89, "closestCanonical": "https://glm52.ai/guides/run-glm-5-2-locally/", "closestIntent": "estimate broad RAM, VRAM, storage, and cost requirements across local deployment routes", "proposedCanonical": "https://glm52.ai/guides/glm-5-2-m5-ultra-mac-studio/", "primaryIntent": "determine whether the announced 512GB M5 Ultra Mac Studio can run GLM-5.2 and whether it is safe to buy for that purpose", "readerJob": "separate checkpoint fit from context headroom and runtime readiness, then define a post-shipment canary before purchase", "distinctIntent": true }, "relatedPageBaseline": { "url": "https://glm52.ai/guides/run-glm-5-2-locally/", "checkedAt": "2026-08-25T15:23:05Z", "httpStatus": 200, "htmlBytes": 122381, "htmlSha256": "55721c22645bc9bd5e36e4987daf8707f63d1f92a9e57fd75ed870e03d62600f", "title": "GLM-5.2 Local Hardware Guide: RAM, VRAM & Cost | GLM52.ai", "description": "Can you run GLM-5.2 locally? See exact BF16, FP8, and quantized model sizes, realistic RAM and GPU needs, 1M-context costs, and when API access wins.", "h1": "Can You Run GLM-5.2 Locally? An Honest Hardware Guide", "canonical": "https://glm52.ai/guides/run-glm-5-2-locally/", "jsonLdCount": 2, "imageCount": 4, "imagesWithAltCount": 4, "noindex": false }, "decision": { "classification": "capacity-qualified-but-purchase-blocked-pending-post-shipment-runtime-canary", "m5Max": "Reject for the listed full-model GGUF artifacts: even UD-IQ2_M exceeds the advertised 128 GB capacity.", "m5Ultra512": "Q4 and lower fit by repository bytes, but Q4 does not have enough conservative gross headroom for the full one-million-token latent-cache estimate, even at one byte per cached value.", "runtime": "Use only a pinned llama.cpp build as the first source-eligible local canary. Stable MLX-LM is blocked, LM Studio lists GLM-5.2 as cloud-only, and no Core AI or Core ML path is published.", "purchase": "Do not preorder or buy the M5 Ultra Mac Studio solely for GLM-5.2 until the 512 GB system ships and the exact quantization passes load, output-quality, context, prompt-processing, decode, thermal, and rollback tests." }, "invariants": { "thirteenHashedSourceReceipts": true, "glm52RevisionPinned": true, "unslothRevisionPinned": true, "m5MaxCannotFitSmallestListedArtifact": true, "q5CannotFitConservativeM5Ultra": true, "q4WeightsFitConservativeM5Ultra": true, "q4FullContextOneByteEstimateDoesNotFit": true, "iq4FullContextTwoByteEstimateFitsBeforeOtherOverhead": true, "llamaCppSupportMerged": true, "mlxStableRejected": true, "lmStudioNotListedForLocalDownload": true, "appleBenchmarkDoesNotNameGlm52": true, "noM5UltraOrModelRun": true, "distinctIntent": true, "relatedPageBaselineIndexable": true } }