{
"schemaVersion": 1,
"operationId": "20260827063900-f8e1483aa0",
"checkedAt": "2026-08-27T07:00:41.352Z",
"status": "pinned-source-zero-model-call-glm52-vllm-thinking-budget-audit",
"revisions": {
"glm52": "b4734de4facf877f85769a911abafc5283eab3d9",
"vllmStable": "v0.28.0",
"vllmStableCommit": "2cf0a6915ce544dc493a0990f2ea38d81601128a",
"vllmRecipe": "5943215a27acb4a243e9d27bd69daf491034cfea",
"glmBudgetIssue": 48201,
"parserIssue": 46040,
"parserFixPr": 45915,
"parserFixMerge": "6c379b9e5439ae305913e4a87ebf2b2e816072b4"
},
"sourceReceipts": [
{
"id": "zai-release",
"url": "https://z.ai/blog/glm-5.2",
"finalUrl": "https://z.ai/blog/glm-5.2",
"purpose": "fixed first-party GLM-5.2 release check",
"httpStatus": 200,
"bytes": 598,
"sha256": "a9e8c2b6f34717d69e3a0aa26bb117256a4d8c95bd299910c2a693000ee88fe8"
},
{
"id": "zhipu-research",
"url": "http://zhipuai.cn/zh/research",
"finalUrl": "https://www.zhipuai.cn/zh/research",
"purpose": "fixed first-party Zhipu research-index check",
"httpStatus": 200,
"bytes": 1235119,
"sha256": "7d50f4c290fbc240f50fabc8d18f7a499688f968f53951655572909ddf547d4e"
},
{
"id": "hf-model-api",
"url": "https://huggingface.co/api/models/zai-org/GLM-5.2",
"finalUrl": "https://huggingface.co/api/models/zai-org/GLM-5.2",
"purpose": "pin the public GLM-5.2 checkpoint revision",
"httpStatus": 200,
"bytes": 23615,
"sha256": "b22ac1d0bc09395ff1b2abb6e439967b0e8c5846e394f6a0d52b8242605a606c"
},
{
"id": "hf-readme",
"url": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/README.md",
"finalUrl": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/README.md",
"purpose": "verify the official vLLM support floor and flexible-effort claim",
"httpStatus": 200,
"bytes": 10905,
"sha256": "ed5aca8ce3dc5f8de626c87e488444343e43b1dcbdeb0e643dc72fea63ab06e8"
},
{
"id": "hf-config",
"url": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/config.json",
"finalUrl": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/config.json",
"purpose": "pin the architecture, vocabulary and native context",
"httpStatus": 200,
"bytes": 3732,
"sha256": "185f93ee6d12548e16a847e279dc0c3c90b1524c970b0866b42fb545747d859a"
},
{
"id": "hf-generation-config",
"url": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/generation_config.json",
"finalUrl": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/generation_config.json",
"purpose": "pin the role-aware EOS token set",
"httpStatus": 200,
"bytes": 194,
"sha256": "ac76b43d8683d3b930126870fc8be73d8679308fe752fa1f381096d8354f6a55"
},
{
"id": "hf-tokenizer",
"url": "https://huggingface.co/zai-org/GLM-5.2/resolve/b4734de4facf877f85769a911abafc5283eab3d9/tokenizer.json",
"finalUrl": "https://huggingface.co/zai-org/GLM-5.2/resolve/b4734de4facf877f85769a911abafc5283eab3d9/tokenizer.json",
"purpose": "verify the think, end-think and stop token IDs",
"httpStatus": 200,
"bytes": 20217442,
"sha256": "19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d"
},
{
"id": "hf-chat-template",
"url": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/chat_template.jinja",
"finalUrl": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/chat_template.jinja",
"purpose": "verify enabled and disabled thinking generation prompts",
"httpStatus": 200,
"bytes": 5076,
"sha256": "172dc74a35e1752df75ecfb2b2cf9326d2852bb1379868ebeec9571654489679"
},
{
"id": "vllm-release-028",
"url": "https://api.github.com/repos/vllm-project/vllm/releases/tags/v0.28.0",
"finalUrl": "https://api.github.com/repos/vllm-project/vllm/releases/tags/v0.28.0",
"purpose": "pin the current audited stable vLLM release",
"httpStatus": 200,
"bytes": 40997,
"sha256": "2f7f99a821df57dcfc0bd66ea60336ed752b611a0203381806ff6d38829e12bb"
},
{
"id": "vllm-reasoning-docs",
"url": "https://raw.githubusercontent.com/vllm-project/vllm/2cf0a6915ce544dc493a0990f2ea38d81601128a/docs/features/reasoning_outputs.md",
"finalUrl": "https://raw.githubusercontent.com/vllm-project/vllm/2cf0a6915ce544dc493a0990f2ea38d81601128a/docs/features/reasoning_outputs.md",
"purpose": "verify the documented enforcement and no-budget behavior",
"httpStatus": 200,
"bytes": 22095,
"sha256": "7431d694e699bf4cc3cf3f83087723882751609980cbbc371348cee4d0b1a7db"
},
{
"id": "vllm-sampling-params",
"url": "https://raw.githubusercontent.com/vllm-project/vllm/2cf0a6915ce544dc493a0990f2ea38d81601128a/vllm/sampling_params.py",
"finalUrl": "https://raw.githubusercontent.com/vllm-project/vllm/2cf0a6915ce544dc493a0990f2ea38d81601128a/vllm/sampling_params.py",
"purpose": "verify accepted and rejected thinking_token_budget values",
"httpStatus": 200,
"bytes": 50759,
"sha256": "73e5bf250cb13803e4ce390407ae059b497898c8b81f9985d2b5c055d4db8f7d"
},
{
"id": "vllm-chat-protocol",
"url": "https://raw.githubusercontent.com/vllm-project/vllm/2cf0a6915ce544dc493a0990f2ea38d81601128a/vllm/entrypoints/openai/chat_completion/protocol.py",
"finalUrl": "https://raw.githubusercontent.com/vllm-project/vllm/2cf0a6915ce544dc493a0990f2ea38d81601128a/vllm/entrypoints/openai/chat_completion/protocol.py",
"purpose": "verify Chat Completions forwards the vLLM-only field",
"httpStatus": 200,
"bytes": 47991,
"sha256": "cb756e3d18e9061a2b306f305e10bd71d43f01ad1b236d0e9cbbb8756cd504dc"
},
{
"id": "vllm-budget-state",
"url": "https://raw.githubusercontent.com/vllm-project/vllm/2cf0a6915ce544dc493a0990f2ea38d81601128a/vllm/v1/sample/thinking_budget_state.py",
"finalUrl": "https://raw.githubusercontent.com/vllm-project/vllm/2cf0a6915ce544dc493a0990f2ea38d81601128a/vllm/v1/sample/thinking_budget_state.py",
"purpose": "audit the token countdown and forced end-marker state",
"httpStatus": 200,
"bytes": 24584,
"sha256": "962f8f55210eb0a431cb9c78b013e35f7a7dd58d06d6fb2e7fcff1b457356f8e"
},
{
"id": "vllm-budget-kernel",
"url": "https://raw.githubusercontent.com/vllm-project/vllm/2cf0a6915ce544dc493a0990f2ea38d81601128a/vllm/v1/worker/gpu/sample/thinking_budget.py",
"finalUrl": "https://raw.githubusercontent.com/vllm-project/vllm/2cf0a6915ce544dc493a0990f2ea38d81601128a/vllm/v1/worker/gpu/sample/thinking_budget.py",
"purpose": "audit the current GPU sampler budget clamp and marker forcing",
"httpStatus": 200,
"bytes": 14998,
"sha256": "020513e1d3406f367de6ad995a95055008fd0c28f2873647e8283c9870e22404"
},
{
"id": "vllm-glm-rfc",
"url": "https://api.github.com/repos/vllm-project/vllm/issues/48201",
"finalUrl": "https://api.github.com/repos/vllm-project/vllm/issues/48201",
"purpose": "capture the attributed GLM-5.2 v0.24.0 budget and tool-call report",
"httpStatus": 200,
"bytes": 6878,
"sha256": "e5c428a3ab1b3610772b6fd51ac52e91fe42f3b4b50257e980898f3b06b49ec5"
},
{
"id": "vllm-parser-issue",
"url": "https://api.github.com/repos/vllm-project/vllm/issues/46040",
"finalUrl": "https://api.github.com/repos/vllm-project/vllm/issues/46040",
"purpose": "capture the GLM reasoning-to-tool parser ordering incident",
"httpStatus": 200,
"bytes": 20015,
"sha256": "3e66630bf95eebe3359dd779882215a977bc77d9a8de0c321538c93d990c50d4"
},
{
"id": "vllm-parser-comments",
"url": "https://api.github.com/repos/vllm-project/vllm/issues/46040/comments?per_page=100",
"finalUrl": "https://api.github.com/repos/vllm-project/vllm/issues/46040/comments?per_page=100",
"purpose": "capture the cross-platform reproduction and merged-fix attribution",
"httpStatus": 200,
"bytes": 7995,
"sha256": "80c93af0d11e9d19c9efe43833741aa1065690d3dd2366301a8f274676fb55ce"
},
{
"id": "vllm-parser-pr",
"url": "https://api.github.com/repos/vllm-project/vllm/pulls/45915",
"finalUrl": "https://api.github.com/repos/vllm-project/vllm/pulls/45915",
"purpose": "pin the merged GLM-5.2 streaming parser engine",
"httpStatus": 200,
"bytes": 19432,
"sha256": "c138026f73d871110abc0779fabf37fa749f04c5e929f024e1b8e278dbda52ad"
},
{
"id": "vllm-recipe",
"url": "https://raw.githubusercontent.com/vllm-project/recipes/5943215a27acb4a243e9d27bd69daf491034cfea/models/zai-org/GLM-5.2.yaml",
"finalUrl": "https://raw.githubusercontent.com/vllm-project/recipes/5943215a27acb4a243e9d27bd69daf491034cfea/models/zai-org/GLM-5.2.yaml",
"purpose": "verify the official glm45 and glm47 launch profile",
"httpStatus": 200,
"bytes": 15247,
"sha256": "517662028453e8a0b62f53a8b507d6b8ac502aaecbd5452b396c37ea2fbae4a7"
},
{
"id": "zai-core-parameters",
"url": "https://docs.z.ai/guides/overview/concept-param",
"finalUrl": "https://docs.z.ai/guides/overview/concept-param",
"purpose": "separate hosted max_tokens and reasoning_effort from the vLLM extension",
"httpStatus": 200,
"bytes": 10742,
"sha256": "f570c786c4ce2bb58f0f16f7809d58328605d711b9a994391457678ac2ffec0e"
},
{
"id": "zai-glm-guide",
"url": "https://docs.z.ai/guides/llm/glm-5.2",
"finalUrl": "https://docs.z.ai/guides/llm/glm-5.2",
"purpose": "verify the hosted GLM-5.2 thinking request surface",
"httpStatus": 200,
"bytes": 25293,
"sha256": "eeca971119068b0619137ce9c9d3c9362131baf16fdab0c067eefe5802919023"
}
],
"modelMarkers": {
"thinkStart": {
"text": "",
"tokenId": 154841
},
"thinkEnd": {
"text": "",
"tokenId": 154842
},
"eosTokenIds": [
154820,
154827,
154829
],
"reportedStopTokenId": 154827,
"enabledGenerationPrefix": "",
"disabledGenerationPrefix": ""
},
"validationMatrix": [
{
"label": "omitted-or-null",
"input": null,
"accepted": true,
"normalized": null,
"meaning": "no explicit reasoning limit",
"actualAccepted": true,
"actualNormalized": null,
"matchesPinnedContract": true
},
{
"label": "minus-one",
"input": -1,
"accepted": true,
"normalized": null,
"meaning": "unlimited sentinel",
"actualAccepted": true,
"actualNormalized": null,
"matchesPinnedContract": true
},
{
"label": "zero",
"input": 0,
"accepted": true,
"normalized": 0,
"meaning": "force the configured reasoning end as soon as the thinking span begins",
"actualAccepted": true,
"actualNormalized": 0,
"matchesPinnedContract": true
},
{
"label": "sixteen",
"input": 16,
"accepted": true,
"normalized": 16,
"meaning": "cap reasoning at 16 tokens",
"actualAccepted": true,
"actualNormalized": 16,
"matchesPinnedContract": true
},
{
"label": "thirty-two",
"input": 32,
"accepted": true,
"normalized": 32,
"meaning": "cap reasoning at 32 tokens",
"actualAccepted": true,
"actualNormalized": 32,
"matchesPinnedContract": true
},
{
"label": "other-negative",
"input": -2,
"accepted": false,
"normalized": null,
"meaning": "rejected",
"actualAccepted": false,
"actualNormalized": null,
"matchesPinnedContract": true
},
{
"label": "float",
"input": 1.5,
"accepted": false,
"normalized": null,
"meaning": "rejected",
"actualAccepted": false,
"actualNormalized": null,
"matchesPinnedContract": true
},
{
"label": "boolean",
"input": true,
"accepted": false,
"normalized": null,
"meaning": "rejected even though Python bool subclasses int",
"actualAccepted": false,
"actualNormalized": null,
"matchesPinnedContract": true
}
],
"capacityFormula": {
"expression": "visible_capacity = max(0, max_tokens - thinking_token_budget - forced_end_marker_tokens - terminal_stop_tokens)",
"forcedEndMarkerTokens": 1,
"terminalStopTokens": 1,
"scope": "A capacity reservation based on the attributed one-token GLM-5.2 end marker and stop token; it does not guarantee answer length, correctness, or a natural stop."
},
"reserveRows": [
{
"maxTokens": 192,
"thinkingTokenBudget": 16,
"minimumVisibleAnswerTokens": 64,
"forcedEndMarkerTokens": 1,
"terminalStopTokens": 1,
"theoreticalVisibleCapacity": 174,
"reserveMargin": 110,
"passesCapacityGate": true
},
{
"maxTokens": 192,
"thinkingTokenBudget": 32,
"minimumVisibleAnswerTokens": 64,
"forcedEndMarkerTokens": 1,
"terminalStopTokens": 1,
"theoreticalVisibleCapacity": 158,
"reserveMargin": 94,
"passesCapacityGate": true
},
{
"maxTokens": 192,
"thinkingTokenBudget": 126,
"minimumVisibleAnswerTokens": 64,
"forcedEndMarkerTokens": 1,
"terminalStopTokens": 1,
"theoreticalVisibleCapacity": 64,
"reserveMargin": 0,
"passesCapacityGate": true
},
{
"maxTokens": 192,
"thinkingTokenBudget": 127,
"minimumVisibleAnswerTokens": 64,
"forcedEndMarkerTokens": 1,
"terminalStopTokens": 1,
"theoreticalVisibleCapacity": 63,
"reserveMargin": -1,
"passesCapacityGate": false
},
{
"maxTokens": 4096,
"thinkingTokenBudget": 512,
"minimumVisibleAnswerTokens": 1024,
"forcedEndMarkerTokens": 1,
"terminalStopTokens": 1,
"theoreticalVisibleCapacity": 3582,
"reserveMargin": 2558,
"passesCapacityGate": true
},
{
"maxTokens": 4096,
"thinkingTokenBudget": 3070,
"minimumVisibleAnswerTokens": 1024,
"forcedEndMarkerTokens": 1,
"terminalStopTokens": 1,
"theoreticalVisibleCapacity": 1024,
"reserveMargin": 0,
"passesCapacityGate": true
},
{
"maxTokens": 4096,
"thinkingTokenBudget": 3071,
"minimumVisibleAnswerTokens": 1024,
"forcedEndMarkerTokens": 1,
"terminalStopTokens": 1,
"theoreticalVisibleCapacity": 1023,
"reserveMargin": -1,
"passesCapacityGate": false
}
],
"attributedIssueRuns": [
{
"label": "unbounded-short-answer",
"source": "vLLM issue #48201",
"maxTokens": 192,
"thinkingTokenBudget": null,
"completionTokens": 192,
"reasoningChars": 344,
"reasoningTokens": null,
"contentChars": 0,
"contentTokens": null,
"finishReason": "length"
},
{
"label": "budget-16-short-answer",
"source": "vLLM issue #48201",
"maxTokens": 192,
"thinkingTokenBudget": 16,
"completionTokens": null,
"reasoningChars": 31,
"reasoningTokens": null,
"contentChars": 287,
"contentTokens": null,
"finishReason": "length"
},
{
"label": "budget-32-token-inspected",
"source": "vLLM issue #48201",
"maxTokens": 192,
"thinkingTokenBudget": 32,
"completionTokens": 87,
"reasoningChars": null,
"reasoningTokens": 32,
"contentChars": null,
"contentTokens": 53,
"endMarkerTokens": 1,
"stopTokens": 1,
"finishReason": "stop"
}
],
"tokenInspectedReconstruction": {
"reasoningTokens": 32,
"endMarkerTokens": 1,
"contentTokens": 53,
"stopTokens": 1,
"reconstructedCompletionTokens": 87,
"reportedCompletionTokens": 87,
"matchesReport": true
},
"versionBoundary": [
{
"version": "v0.23.0",
"evidence": "official GLM-5.2 model card and vLLM recipe set the general serving floor",
"glmBudgetStatus": "not established by this audit"
},
{
"version": "v0.24.0",
"evidence": "open vLLM issue #48201 reports a successful GLM-5.2 budget and streaming tool-call run",
"glmBudgetStatus": "reporter-observed, not a merged GLM-specific regression"
},
{
"version": "v0.28.0",
"evidence": "tagged source accepts, forwards and enforces the generic field; current E2E tests cover other reasoning models",
"glmBudgetStatus": "source-audited generic implementation; GLM-specific RFC remains open"
}
],
"overlapAudit": {
"prepublicationCanonicalCount": 92,
"registeredContentPages": 85,
"exactExistingFieldMentions": 0,
"nearestPages": [
{
"url": "https://glm52.ai/guides/glm-5-2-max-tokens/",
"intent": "set the total output cap and fit input plus output within the effective context",
"difference": "does not configure or validate the vLLM-only reasoning sub-budget"
},
{
"url": "https://glm52.ai/guides/glm-5-2-reasoning-effort/",
"intent": "choose Z.AI hosted reasoning-effort labels",
"difference": "does not enforce a numeric self-hosted reasoning-token ceiling"
},
{
"url": "https://glm52.ai/guides/glm-5-2-tool-calling/",
"intent": "implement the hosted end-to-end function loop",
"difference": "does not validate the local vLLM reasoning-to-tool parser transition"
}
],
"distinctIntent": true
},
"searchSupply": {
"method": "one exact Google query through the budgeted SerpAPI client",
"query": "GLM-5.2 vLLM thinking_token_budget",
"retrievedAt": "2026-08-27T06:50:32Z",
"organicResultCount": 9,
"approximateGoogleResultCount": 576,
"exactSpelling": true,
"dedicatedGuideFound": false,
"notableResults": [
"official vLLM GLM-5.2 recipe",
"Z.AI GLM-5.2 release",
"vLLM GLM parser issue #46040"
],
"serpApiRequests": 1,
"evaluation": "high-value",
"decisionChanged": true,
"caveat": "The approximate result count is not search volume and is not used as a demand claim."
},
"runtimeBoundary": {
"modelCalls": 0,
"localGpuRuns": 0,
"upstreamTestRuns": 0,
"dockerRuns": 0,
"paidSerpApiRequests": 1,
"sourceClone": "A task-owned shallow vLLM v0.28.0 checkout was inspected and removed by a trap.",
"limitation": "Issue #48201 supplies attributed runtime observations; this audit does not reproduce GLM-5.2 generation or tool calls."
},
"aiHot": [
{
"itemId": "cmtb43lik0i9sroamwtuxb1bu",
"classification": "duplicate-intent",
"note": "GLM-5.3 Flash routing overlaps the existing GLM-5.2 versus GLM-5.3 comparison and API migration pages; it does not create a new GLM-5.2 reader job."
},
{
"itemId": "cmtas790607swroamcmunq53y",
"classification": "weak-glm-link",
"note": "NVIDIA revenue guidance is infrastructure-market news and offers no GLM-5.2 configuration or model decision."
},
{
"itemId": "cmtar8e770746roam1wznm4mi",
"classification": "weak-glm-link",
"note": "The AWS GPU order does not verify a GLM-5.2 access, cost, compatibility, or deployment choice."
},
{
"itemId": "cmtaq20ih062wroamdrhb54b6",
"classification": "weak-glm-link",
"note": "NVIDIA earnings are not a direct GLM-5.2 workload or operating decision."
},
{
"itemId": "cmtaj3732034urovudls0byfe",
"classification": "weak-glm-link",
"note": "GlucoFM is a glucose-monitoring model with no direct GLM-5.2 alternative or integration."
},
{
"itemId": "cmtaej1vz0czhroj2aybdbq26",
"classification": "weak-glm-link",
"note": "Claude in Chrome is a separate proprietary browser agent and does not supply a concrete GLM-5.2 decision beyond a bolted-on comparison."
},
{
"itemId": "cmtaddncd0aq0roj2xmay0scj",
"classification": "weak-glm-link",
"note": "Anthropic's research-data program does not create a GLM-5.2 access, evaluation, or deployment task."
},
{
"itemId": "cmtaxja5w0ckaroamws5is8wp",
"classification": "weak-glm-link",
"note": "NVIDIA quarterly guidance is general market news without a GLM-5.2-specific reader outcome."
},
{
"itemId": "cmtaighmj02k5rovu2z28rxc2",
"classification": "duplicate-intent",
"note": "The Hugging Face incident already has a dedicated GLM-5.2 forensics canonical; another event wrapper would duplicate its verification job."
}
],
"decision": {
"classification": "use-a-bounded-vllm-only-reasoning-budget-with-an-explicit-answer-reserve-and-parser-canary",
"fieldBoundary": "`thinking_token_budget` is a vLLM sampling extension in this audit, not a documented Z.AI hosted parameter.",
"requestRule": "Use a non-negative integer; -1 or omission means unlimited, while booleans, floats, and other negative values must fail.",
"capacityRule": "Choose the visible-answer reserve first, subtract the one-token GLM-5.2 reasoning end marker and one terminal stop token, then cap reasoning with the remaining allowance.",
"parserRule": "For tool workloads, require reasoning chunks to end before tool-call chunks, valid JSON arguments, finish_reason=tool_calls, and no marker text inside arguments.",
"releaseRule": "Pin vLLM source and image. The official general serving floor is v0.23.0, one report observes the field on v0.24.0, and v0.28.0 contains the audited generic implementation; none of those facts is a GLM-specific production guarantee.",
"fallback": "Disable thinking for short deterministic tool dispatch when that preserves task acceptance; otherwise raise max_tokens or lower the reasoning budget only after a frozen acceptance test."
},
"invariants": {
"twentyOneSourceReceipts": true,
"allSourcesReturned200": true,
"allSourceHashesValid": true,
"tokenizerArtifactWasResolved": true,
"pinnedMarkersMatch": true,
"validationCasesMatchPinnedContract": true,
"boundaryCapacityCellsFlip": true,
"issueTokenLayoutReconstructed": true,
"versionBoundaryIsQualified": true,
"distinctIntent": true,
"onePaidSerpRequest": true,
"allAiHotClassified": true,
"sevenWeakLinksAndTwoDuplicates": true,
"noModelOrGpuRun": true
}
}