{ "schemaVersion": "horizon.agent-claims.v1", "project": "Horizon RunMap", "reviewedDate": "2026-09-16", "projectType": "experimental MoE execution architecture", "author": "Lucas Ribeiro", "runtimeSourceAccess": "private by author decision", "primaryQuestion": "How can storage precision, execution precision, residency and data movement be coordinated under a constrained GPU-memory budget?", "architecture": "/architecture.json", "technicalBrief": "/technical-brief.md", "claims": [ { "id": "selective-compression", "status": "implemented-in-recorded-paths", "statement": "Selective INT4 compression of routed experts with a separate FP16 non-routed path", "source": "/architecture.md" }, { "id": "compact-execution", "status": "implemented-in-recorded-paths", "statement": "Compact expert GPU execution with FP16 activations and recorded scalar/grouped executors", "source": "/evidence.md" }, { "id": "partial-residency", "status": "measured", "experimentId": "residency39", "statement": "Partial residency observations under the selected campaign contract" }, { "id": "full-residency", "status": "measured", "experimentId": "residency39", "statement": "OLMoE c64 full residency observations with Grouped T3" }, { "id": "instruction-following", "status": "measured", "experimentId": "ifeval541", "statement": "Complete IFEval instruction-following evaluation of retained Horizon responses" } ], "notClaimed": [ "fastest MoE runtime", "SOTA throughput", "runtime-to-runtime superiority", "superiority over llama.cpp", "general BF16 quality equivalence", "validated frontier-scale deployment", "independently reproduced runtime generation", "world-first quantization, offloading or caching" ], "experimentalWarnings": { "olmoeExecutor": "OLMoE c30-c60 use PRODUCT-core scalar; c64 uses Grouped T3. This is not a residency-only comparison.", "qualityPopulations": "IFEval 541 and the earlier 120-pair behavioral study are separate experiments.", "reproduction": "Published artifacts enable their declared verification, not full independent runtime reproduction.", "architecture": "A shared architectural explanation does not certify every executor against every ADR 0009 invariant." }, "experiments": [ { "id": "residency39", "editionId": "technical-preview-residency39-v3", "manifest": "/evidence/releases/technical-preview-residency39-v3/manifest.json", "manifestSha256": "87b957c4e06be8bf453974959c1017f7f0240bdf45b0d430463f9a022014a93b", "validationStatus": "valid", "evidenceLevel": "E3", "publicationStatus": "public", "reproductionStatus": "not_attempted", "experimentType": "controlled_research", "configurations": 13, "workers": 39, "measuredResponses": 468, "promptsPerBlock": 12, "blocks": 3, "tokensPerResponse": 128, "contract": "PERF-REAL-v1; 12 prompts repeated in 3 blocks; 36 measured responses of 128 tokens per configuration; greedy selection, EOS suppressed; 64-token warmup excluded. Canonical initial placement, evolving LRU through fixed prompts; concurrency 1.", "verificationScope": "hash and curated aggregate consistency; no complete raw-run recomputation", "results": [ { "id": "olmoe-0924-instruct-c30-residency39", "model": "allenai/OLMoE-1B-7B-0924-Instruct", "revision": "7f1c97f440f06ce36705e4f2b843edb5925f4498", "source": "/evidence/releases/technical-preview-residency39-v3/olmoe-0924-instruct-c30-residency39.json", "executor": "PRODUCT-core scalar", "residentExpertsPerLayer": 30, "totalExpertsPerLayer": 64, "decodeTokensPerSecond": 23.001858869666933, "aggregation": "pooled-post-first-token", "comparisonGroup": "olmoe0924-scalar-partial", "ttftSeconds": 6.170728905555555, "responseSeconds": 11.692021802777779, "wholeWorkerPeakVramGiB": 4.113231658935547, "wholeWorkerPeakRssGiB": 4.946601867675781, "measuredRequestH2dGiB": 1108.095474243164, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "olmoe-0924-instruct-c40-residency39", "model": "allenai/OLMoE-1B-7B-0924-Instruct", "revision": "7f1c97f440f06ce36705e4f2b843edb5925f4498", "source": "/evidence/releases/technical-preview-residency39-v3/olmoe-0924-instruct-c40-residency39.json", "executor": "PRODUCT-core scalar", "residentExpertsPerLayer": 40, "totalExpertsPerLayer": 64, "decodeTokensPerSecond": 27.691961601492615, "aggregation": "pooled-post-first-token", "comparisonGroup": "olmoe0924-scalar-partial", "ttftSeconds": 4.486795727777778, "responseSeconds": 9.072964155555557, "wholeWorkerPeakVramGiB": 4.618438720703125, "wholeWorkerPeakRssGiB": 4.950050354003906, "measuredRequestH2dGiB": 576.1717987060547, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "olmoe-0924-instruct-c50-residency39", "model": "allenai/OLMoE-1B-7B-0924-Instruct", "revision": "7f1c97f440f06ce36705e4f2b843edb5925f4498", "source": "/evidence/releases/technical-preview-residency39-v3/olmoe-0924-instruct-c50-residency39.json", "executor": "PRODUCT-core scalar", "residentExpertsPerLayer": 50, "totalExpertsPerLayer": 64, "decodeTokensPerSecond": 31.501747677273897, "aggregation": "pooled-post-first-token", "comparisonGroup": "olmoe0924-scalar-partial", "ttftSeconds": 3.6184148333333335, "responseSeconds": 7.649937188888888, "wholeWorkerPeakVramGiB": 5.147003173828125, "wholeWorkerPeakRssGiB": 4.950908660888672, "measuredRequestH2dGiB": 230.88111877441406, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "olmoe-0924-instruct-c60-residency39", "model": "allenai/OLMoE-1B-7B-0924-Instruct", "revision": "7f1c97f440f06ce36705e4f2b843edb5925f4498", "source": "/evidence/releases/technical-preview-residency39-v3/olmoe-0924-instruct-c60-residency39.json", "executor": "PRODUCT-core scalar", "residentExpertsPerLayer": 60, "totalExpertsPerLayer": 64, "decodeTokensPerSecond": 35.01159000794033, "aggregation": "pooled-post-first-token", "comparisonGroup": "olmoe0924-scalar-partial", "ttftSeconds": 3.0013486361111115, "responseSeconds": 6.628718886111112, "wholeWorkerPeakVramGiB": 5.670684814453125, "wholeWorkerPeakRssGiB": 4.956043243408203, "measuredRequestH2dGiB": 37.615814208984375, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "olmoe-0924-instruct-c64-residency39", "model": "allenai/OLMoE-1B-7B-0924-Instruct", "revision": "7f1c97f440f06ce36705e4f2b843edb5925f4498", "source": "/evidence/releases/technical-preview-residency39-v3/olmoe-0924-instruct-c64-residency39.json", "executor": "Grouped T3", "residentExpertsPerLayer": 64, "totalExpertsPerLayer": 64, "decodeTokensPerSecond": 61.25592960111646, "aggregation": "pooled-post-first-token", "comparisonGroup": null, "ttftSeconds": 0.7455810444444444, "responseSeconds": 2.818849719444444, "wholeWorkerPeakVramGiB": 5.990997314453125, "wholeWorkerPeakRssGiB": 4.983863830566406, "measuredRequestH2dGiB": 0, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "olmoe-0125-instruct-c30-residency39", "model": "allenai/OLMoE-1B-7B-0125-Instruct", "revision": "b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e", "source": "/evidence/releases/technical-preview-residency39-v3/olmoe-0125-instruct-c30-residency39.json", "executor": "PRODUCT-core scalar", "residentExpertsPerLayer": 30, "totalExpertsPerLayer": 64, "decodeTokensPerSecond": 23.04059913975746, "aggregation": "pooled-post-first-token", "comparisonGroup": "olmoe0125-scalar-partial", "ttftSeconds": 6.070801397222223, "responseSeconds": 11.58281083888889, "wholeWorkerPeakVramGiB": 4.148338317871094, "wholeWorkerPeakRssGiB": 4.968273162841797, "measuredRequestH2dGiB": 1115.870361328125, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "olmoe-0125-instruct-c40-residency39", "model": "allenai/OLMoE-1B-7B-0125-Instruct", "revision": "b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e", "source": "/evidence/releases/technical-preview-residency39-v3/olmoe-0125-instruct-c40-residency39.json", "executor": "PRODUCT-core scalar", "residentExpertsPerLayer": 40, "totalExpertsPerLayer": 64, "decodeTokensPerSecond": 27.12311586468119, "aggregation": "pooled-post-first-token", "comparisonGroup": "olmoe0125-scalar-partial", "ttftSeconds": 4.526311008333334, "responseSeconds": 9.2086639, "wholeWorkerPeakVramGiB": 4.615997314453125, "wholeWorkerPeakRssGiB": 4.951240539550781, "measuredRequestH2dGiB": 580.71533203125, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "olmoe-0125-instruct-c50-residency39", "model": "allenai/OLMoE-1B-7B-0125-Instruct", "revision": "b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e", "source": "/evidence/releases/technical-preview-residency39-v3/olmoe-0125-instruct-c50-residency39.json", "executor": "PRODUCT-core scalar", "residentExpertsPerLayer": 50, "totalExpertsPerLayer": 64, "decodeTokensPerSecond": 31.22055391154392, "aggregation": "pooled-post-first-token", "comparisonGroup": "olmoe0125-scalar-partial", "ttftSeconds": 3.5939765083333333, "responseSeconds": 7.661809525, "wholeWorkerPeakVramGiB": 5.150886535644531, "wholeWorkerPeakRssGiB": 4.954097747802734, "measuredRequestH2dGiB": 235.35324096679688, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "olmoe-0125-instruct-c60-residency39", "model": "allenai/OLMoE-1B-7B-0125-Instruct", "revision": "b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e", "source": "/evidence/releases/technical-preview-residency39-v3/olmoe-0125-instruct-c60-residency39.json", "executor": "PRODUCT-core scalar", "residentExpertsPerLayer": 60, "totalExpertsPerLayer": 64, "decodeTokensPerSecond": 35.442592698091474, "aggregation": "pooled-post-first-token", "comparisonGroup": "olmoe0125-scalar-partial", "ttftSeconds": 2.9682318777777774, "responseSeconds": 6.551491180555555, "wholeWorkerPeakVramGiB": 5.670722961425781, "wholeWorkerPeakRssGiB": 4.959484100341797, "measuredRequestH2dGiB": 40.81146240234375, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "olmoe-0125-instruct-c64-residency39", "model": "allenai/OLMoE-1B-7B-0125-Instruct", "revision": "b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e", "source": "/evidence/releases/technical-preview-residency39-v3/olmoe-0125-instruct-c64-residency39.json", "executor": "Grouped T3", "residentExpertsPerLayer": 64, "totalExpertsPerLayer": 64, "decodeTokensPerSecond": 62.07316072245486, "aggregation": "pooled-post-first-token", "comparisonGroup": null, "ttftSeconds": 0.6966347027777778, "responseSeconds": 2.7426075277777775, "wholeWorkerPeakVramGiB": 6.010833740234375, "wholeWorkerPeakRssGiB": 4.962856292724609, "measuredRequestH2dGiB": 0, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "qwen15-moe-chat-c20-residency39", "model": "Qwen/Qwen1.5-MoE-A2.7B-Chat", "revision": "ec052fda178e241c7c443468d2fa1db6618996be", "source": "/evidence/releases/technical-preview-residency39-v3/qwen15-moe-chat-c20-residency39.json", "executor": "top4-fused-v1", "residentExpertsPerLayer": 20, "totalExpertsPerLayer": 60, "decodeTokensPerSecond": 15.22236659827508, "aggregation": "pooled-post-first-token", "comparisonGroup": "qwen-top4", "ttftSeconds": 9.41710652777778, "responseSeconds": 17.760093091666665, "wholeWorkerPeakVramGiB": 7.110137939453125, "wholeWorkerPeakRssGiB": 7.967433929443359, "measuredRequestH2dGiB": 2702.7495861053467, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "qwen15-moe-chat-c30-residency39", "model": "Qwen/Qwen1.5-MoE-A2.7B-Chat", "revision": "ec052fda178e241c7c443468d2fa1db6618996be", "source": "/evidence/releases/technical-preview-residency39-v3/qwen15-moe-chat-c30-residency39.json", "executor": "top4-fused-v1", "residentExpertsPerLayer": 30, "totalExpertsPerLayer": 60, "decodeTokensPerSecond": 17.170567206372947, "aggregation": "pooled-post-first-token", "comparisonGroup": "qwen-top4", "ttftSeconds": 7.506468494444444, "responseSeconds": 14.902846172222224, "wholeWorkerPeakVramGiB": 8.077144622802734, "wholeWorkerPeakRssGiB": 7.992816925048828, "measuredRequestH2dGiB": 1852.0655822753906, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] }, { "id": "qwen15-moe-chat-c39-residency39", "model": "Qwen/Qwen1.5-MoE-A2.7B-Chat", "revision": "ec052fda178e241c7c443468d2fa1db6618996be", "source": "/evidence/releases/technical-preview-residency39-v3/qwen15-moe-chat-c39-residency39.json", "executor": "top4-fused-v1", "residentExpertsPerLayer": 39, "totalExpertsPerLayer": 60, "decodeTokensPerSecond": 19.945904586527387, "aggregation": "pooled-post-first-token", "comparisonGroup": "qwen-top4", "ttftSeconds": 5.799750786111111, "responseSeconds": 12.16697266111111, "wholeWorkerPeakVramGiB": 8.989349365234375, "wholeWorkerPeakRssGiB": 7.973819732666016, "measuredRequestH2dGiB": 1172.502737045288, "limitations": [ "Controlled E3 measurement on a shared desktop under the recorded configuration.", "Driver VRAM and process RSS are whole-worker peaks including preparation and warmup; VRAM includes desktop use.", "OLMoE c60 to c64 changes executor as well as residency; the transition is not a residency-only effect.", "Block spread is an observed range, not a confidence interval. Token parity concerns repetitions within each configuration only.", "TTFT and response time use internal request boundaries; isolated prefill is UNKNOWN and OLMoE HTTP observation is UNSUPPORTED." ] } ] }, { "id": "behavioral120", "editionId": "olmoe-0924-fidelity-v2", "manifest": "/evidence/behavioral/olmoe-0924-fidelity-v2/manifest.json", "manifestSha256": "8105c554260e54974990560cfeca65b9302e9c7480663dad2a0bf1d8c8a939b2", "validationStatus": "valid", "evidenceLevel": "E3", "publicationStatus": "public", "reproductionStatus": "not_attempted", "experimentType": "paired_structural_behavior", "promptCount": 120, "metrics": { "both_contract_valid_cases": 21, "category_counts": { "exact-token-parity": 16, "horizon-contract-improvement": 7, "horizon-contract-regression": 8, "non-directional-divergence": 82, "semantic-divergence": 4, "semantic-json-parity": 3 }, "contract_sha256": "867d6498f374200ac1cf7566aeaa5c492d2f8b253f8048bf48594171a8fa591b", "directionless_divergences": 86, "exact_token_parity": 16, "horizon_contract_improvements": 7, "horizon_contract_regressions": 8, "horizon_contract_valid_cases": 28, "horizon_regression_eligible_cases": 29, "native_contract_valid_cases": 29, "observable_behavior_parity": 19, "prompt_count": 120, "schema_version": "horizon.olmoe-0924.behavioral-fidelity.summary.v2", "semantic_json_parity": 3 }, "contract": "/evidence/behavioral/olmoe-0924-fidelity-v2/contract.json", "verificationScope": "CPU recomputation of frozen classifications from retained response pairs; no answer-correctness grade or generation rerun" }, { "id": "ifeval541", "editionId": "olmoe-0924-ifeval541-v1", "manifest": "/evidence/ifeval/olmoe-0924-ifeval541-v1/manifest.json", "manifestSha256": "9dcdb178a5469ca57c74c1e72ed2956cc94539b16607fdee7771b3b5442c8ed2", "validationStatus": "valid", "evidenceLevel": "E1", "experimentType": "standalone_benchmark", "publicationStatus": "public", "reproductionStatus": "not_attempted", "model": "allenai/OLMoE-1B-7B-0924-Instruct", "revision": "7f1c97f440f06ce36705e4f2b843edb5925f4498", "metrics": { "inst_level_loose_acc": { "decimal": 0.5683453237410072, "denominator": 834, "numerator": 474 }, "inst_level_strict_acc": { "decimal": 0.5191846522781774, "denominator": 834, "numerator": 433 }, "prompt_level_loose_acc": { "decimal": 0.4565619223659889, "denominator": 541, "numerator": 247 }, "prompt_level_strict_acc": { "decimal": 0.4011090573012939, "denominator": 541, "numerator": 217 }, "schema": "horizon.ifeval-541.summary.v1" }, "termination": { "eos": 467, "length": 74 }, "contract": "/evidence/ifeval/olmoe-0924-ifeval541-v1/observations/contract.json", "verificationScope": "CPU recomputation using frozen dataset and scorer over retained responses; no generation rerun", "externalReferences": [ { "value": 48.1, "label": "Original 0924 model card", "metric": "Loose Acc", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0924-Instruct#evaluation-snapshot" }, { "value": 45.29, "label": "0924 checkpoint in the 0125 model card", "metric": "IFEval (variant unspecified in table)", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct#performance" } ], "externalReferenceRole": "contextual, non-controlled", "preservationClaimAllowed": false } ] }