rdrr / registry /catalog.json
clocksmith's picture
registry: use lowercase clocksmith namespace
3968b13 verified
Raw History Blame Contribute Delete
51.7 kB
{
"version": 1,
"lifecycleSchemaVersion": 1,
"updatedAt": "2026-07-19",
"models": [
{
"modelId": "gemma-3-270m-it-q4k-ehf16-af32",
"family": "gemma3",
"classification": {
"domain": "language",
"tasks": [
"generation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text"
],
"outputs": [
"text"
]
},
"label": "Gemma 3 270M Instruct (Q4K/F32a)",
"description": "Gemma 3 270M text model. Q4K weights, F16 embeddings/head, F32 accumulation.",
"modes": [
"text"
],
"aliases": [
"gemma3-270m",
"google/gemma-3-270m-it",
"gemma-3-270m-it-wq4k-ef16",
"gemma-3-270m-it-wq4k-ef16-hf16",
"gemma-3-270m-it-wq4k-ef16-hf16-f32"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "a8591b20bce7c22d75becde1315482e76ff85fc9",
"path": "models/gemma-3-270m-it-q4k-ehf16-af32"
},
"baseUrl": null,
"quickstart": true,
"demoVisible": true,
"demoLabel": "Gemma 3 270M",
"demoRole": "Stable F32 browser lane",
"demoSortOrder": 10,
"sizeBytes": 418518521,
"recommended": false,
"sortOrder": 10,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "inference",
"surface": [
"browser",
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-06-24",
"source": "runtime-verify",
"notes": "Browser/WebGPU quickstart smoke is package-visible at reports/release-claims/gemma-3-270m-it-q4k-ehf16-af32/2026-06-24T01-11-24.275Z.json. The report produced a coherent sky-color answer, stopReason=stop-token, executionContractOk=true, and decodeTokensPerSec > 0.",
"contracts": {
"executionContractOk": true
}
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/gemma-3-270m-it-ONNX",
"dtype": "q4f16"
}
},
"benchmarkEvidence": {
"status": "benchmark-selected",
"localClaimLaneId": "gemma-3-270m-it-q4k-rdrr",
"runtimeReport": "reports/release-claims/gemma-3-270m-it-q4k-ehf16-af32/2026-06-24T01-11-24.275Z.json",
"compareResult": "benchmarks/vendors/results/compare_20260627T200811.json",
"summarySvg": "benchmarks/vendors/results/gemma-3-270m-it-q4k-rdrr-browser-p512-d128-t0-k1-strix-halo-20260627T200811.svg"
},
"sourceCheckpointId": "google/gemma-3-270m-it",
"weightPackId": "gemma-3-270m-it-q4k-ehf16-af32-wp-catalog-v1",
"manifestVariantId": "gemma-3-270m-it-q4k-ehf16-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "google-embeddinggemma-300m-q4k-ehf16-af32",
"family": "embeddinggemma",
"classification": {
"domain": "language",
"tasks": [
"embedding"
],
"architectureRole": "decoder-embedding-encoder",
"inputs": [
"text"
],
"outputs": [
"pooled-embedding"
]
},
"label": "EmbeddingGemma 300M (Q4K)",
"description": "EmbeddingGemma 300M sentence embedding model. Q4K weights, F16 embeddings/head, F32 accumulation.",
"modes": [
"embedding"
],
"aliases": [
"embeddinggemma-300m",
"google/embeddinggemma-300m",
"google-embeddinggemma-300m-wq4k-ef16"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "95f0b29ec73dea70394c6bcfa8407bc6796df6c9",
"path": "models/google-embeddinggemma-300m-q4k-ehf16-af32"
},
"baseUrl": null,
"quickstart": true,
"sizeBytes": 480563695,
"recommended": true,
"sortOrder": 11,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "embedding",
"surface": [
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-06-24",
"source": "runtime-verify",
"contracts": {
"executionContractOk": true
},
"notes": "Node/WebGPU embedding verification is package-visible at reports/release-claims/google-embeddinggemma-300m-q4k-ehf16-af32/2026-06-24T01-42-03.814Z.json. The report produced finite 768-dimensional unit-norm embeddings and passed the semantic retrieval/pair suite thresholds. Registry verification uses profiles/vector-stability so the manifest-owned f32 KV-cache contract is preserved."
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/embeddinggemma-300m-ONNX",
"dtype": "q4"
}
},
"sourceCheckpointId": "google/embeddinggemma-300m",
"weightPackId": "google-embeddinggemma-300m-q4k-ehf16-af32-wp-catalog-v1",
"manifestVariantId": "google-embeddinggemma-300m-q4k-ehf16-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "gemma-3-1b-it-q4k-ehf16-af32",
"family": "gemma3",
"classification": {
"domain": "language",
"tasks": [
"generation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text"
],
"outputs": [
"text"
]
},
"label": "Gemma 3 1B Instruct (Q4K/F32a)",
"description": "Gemma 3 1B text model. Q4K weights, F16 embeddings/head, F32 accumulation.",
"modes": [
"text"
],
"aliases": [
"gemma3-1b",
"google/gemma-3-1b-it",
"gemma-3-1b-it-wq4k-ef16-hf16"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "7c3d30e300bcb02cbd68fb0db3eee64fbf738f99",
"path": "models/gemma-3-1b-it-q4k-ehf16-af32"
},
"baseUrl": null,
"quickstart": true,
"demoLabel": "Gemma 3 1B",
"demoRole": "Larger Gemma text model",
"sizeBytes": 1044786820,
"recommended": false,
"sortOrder": 12,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "inference",
"surface": [
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-06-24",
"source": "runtime-verify",
"notes": "Node/WebGPU quickstart smoke is package-visible at reports/release-claims/gemma-3-1b-it-q4k-ehf16-af32/2026-06-24T01-23-10.494Z.json. The report produced coherent text, executionContractOk=true, and decodeTokensPerSec > 0.",
"contracts": {
"executionContractOk": true
}
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/gemma-3-1b-it-ONNX-GQA",
"dtype": "q4f16"
}
},
"sourceCheckpointId": "google/gemma-3-1b-it",
"weightPackId": "gemma-3-1b-it-q4k-ehf16-af32-wp-catalog-v1",
"manifestVariantId": "gemma-3-1b-it-q4k-ehf16-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "gemma-4-e2b-it-q4k-ehf16-af32",
"family": "gemma4",
"classification": {
"domain": "language",
"tasks": [
"generation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text",
"image",
"audio"
],
"outputs": [
"text"
]
},
"label": "Gemma 4 E2B Instruct (Q4K/F32a)",
"description": "Gemma 4 E2B multimodal model. Q4K weights, F16 embeddings/head/vision/audio/projector, F32 accumulation.",
"modes": [
"text",
"vision"
],
"aliases": [
"gemma4-e2b",
"google/gemma-4-e2b-it"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "2070777c77047d54e6eae105f6dcb1891cf6f21a",
"path": "models/gemma-4-e2b-it-q4k-ehf16-af32"
},
"baseUrl": null,
"quickstart": true,
"demoVisible": false,
"sizeBytes": 7094076955,
"recommended": false,
"sortOrder": 13,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "inference",
"surface": [
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-04-18",
"source": "runtime-verify",
"notes": "Package-visible runtime evidence is staged at reports/release-claims/gemma-4-e2b-it-q4k-ehf16-af32/2026-04-18T19-17-45.997Z.json. Browser and Electron remain unverified in the Gemma 4 target matrix until same-surface runtime receipts are promoted.",
"contracts": {
"executionContractOk": true
}
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/gemma-4-E2B-it-ONNX",
"dtype": "q4f16"
}
},
"sourceCheckpointId": "google/gemma-4-E2B-it",
"weightPackId": "gemma-4-e2b-it-q4k-ehf16-af32-wp-catalog-v1",
"manifestVariantId": "gemma-4-e2b-it-q4k-ehf16-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
"family": "gemma4",
"classification": {
"domain": "language",
"tasks": [
"generation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text",
"image",
"audio"
],
"outputs": [
"text"
]
},
"label": "Gemma 4 E2B Instruct (Q4K/F32a/INT4 PLE)",
"description": "Gemma 4 E2B multimodal model. Q4K weights, INT4 per-layer embeddings with per-row F32 scales, F16 embeddings/head/vision/audio/projector, F32 accumulation.",
"modes": [
"text",
"vision"
],
"aliases": [
"gemma4-e2b-int4ple",
"gemma-4-e2b-int4ple"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"path": "models/gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
"revision": "16256bf16dc8f92b8fe5105b07628cef91587f0a"
},
"baseUrl": null,
"quickstart": true,
"demoLabel": "Gemma 4 E2B",
"demoRole": "Gemma 4 text and vision",
"demoSortOrder": 14,
"demoPreferredVariantId": "gemma-4-e2b-it-q4k-ehf16-af16-int4ple",
"sizeBytes": 4009916310,
"recommended": false,
"sortOrder": 14,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "inference",
"surface": [
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-04-22",
"source": "runtime-verify",
"notes": "Committed program-bundle reference reports/program-bundles/gemma-4-e2b-it-q4k-ehf16-af32-int4ple/2026-04-22T18-09-10.471Z.reference.json decodes the sky prompt to blue with executionContractOk=true. Browser and Electron remain unverified in the Gemma 4 target matrix until same-surface runtime receipts are promoted.",
"contracts": {
"executionContractOk": true,
"layerPatternContractOk": true
}
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/gemma-4-E2B-it-ONNX",
"dtype": "q4f16"
}
},
"sourceCheckpointId": "google/gemma-4-E2B-it",
"weightPackId": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple-wp-catalog-v1",
"manifestVariantId": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "translategemma-4b-1b-enes-q4k-ehf16-af32",
"family": "translategemma",
"classification": {
"domain": "language",
"tasks": [
"translation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text"
],
"outputs": [
"text"
]
},
"label": "TranslateGemma 1B EN/ES Student (NativeKD2 Q4K)",
"description": "Experimental-tier English/Spanish translation specialist distilled from TranslateGemma 4B into Gemma 3 1B. The hosted Q4K artifact has deterministic WMT13 and browser/WebGPU evidence; quality and performance claims remain limited to the recorded EN/ES and Radeon/Vulkan receipts.",
"modes": [
"translate"
],
"aliases": [
"translategemma-enes-1b",
"translategemma-nativekd2-enes-1b",
"translategemma-1b-translation-student"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"path": "models/translategemma-4b-1b-enes-q4k-ehf16-af32",
"revision": "1e2c047357f07dfe45a8effc2fa915acd73999ee"
},
"baseUrl": null,
"quickstart": false,
"demoVisible": true,
"demoLabel": "TranslateGemma EN/ES",
"demoRole": "English and Spanish translation",
"sizeBytes": 1025493248,
"recommended": false,
"sortOrder": 14.9,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "compare",
"tested": "verified"
},
"tested": {
"suite": "translation-quality",
"surface": [
"browser",
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-07-11",
"source": "runtime-verify",
"notes": "The exact Q4K artifact passed shard integrity, deterministic EN/ES smoke outputs, 8/8 repeated-token checks, and a 128-row WMT13 evaluation at 31.9149 BLEU / 58.2124 chrF. Baseline and optimized Doppler outputs match 128/128. The pinned hosted revision also passed browser/WebGPU verification and a 2+3-run optimized benchmark with exact expected output. Experimental visibility preserves the measured 2.2748 BLEU / 1.7264 chrF Q4K gap to the source BF16 checkpoint and does not imply Apple Metal evidence.",
"contracts": {
"executionContractOk": true
}
}
},
"vendorBenchmark": null,
"sourceCheckpointId": "clocksmith/gamma-translategemma-4b-1b-enes-nativekd2@sha256:c567331d8eb9d3da65205b83f4e0af06568c47ba81a7927005c1c2e72444bc87",
"weightPackId": "translategemma-4b-1b-enes-q4k-ehf16-af32-wp-16acf07340bd",
"manifestVariantId": "translategemma-4b-1b-enes-q4k-ehf16-af32-mv-b3c719ee90e7",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "gemma-4-e2b-it-q4k-ehf16-af16-int4ple",
"family": "gemma4",
"classification": {
"domain": "language",
"tasks": [
"generation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text",
"image",
"audio"
],
"outputs": [
"text"
]
},
"label": "Gemma 4 E2B Instruct (Q4K/F16a/INT4 PLE)",
"description": "Gemma 4 E2B multimodal Doppler/WebGPU f16 activation variant. Shares the existing Q4K/INT4 PLE weight pack and falls back to AF32 when capability rules reject AF16.",
"modes": [
"text",
"vision"
],
"aliases": [
"gemma4-e2b-int4ple-f16",
"gemma4-e2b-af16-int4ple",
"gemma-4-e2b-af16-int4ple"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "2d326649ae3d04b83323de0e6247e1e42fb1d9af",
"path": "models/gemma-4-e2b-it-q4k-ehf16-af16-int4ple"
},
"baseUrl": null,
"quickstart": false,
"demoVisible": false,
"sizeBytes": 4009916310,
"recommended": false,
"sortOrder": 15,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "inference",
"surface": [
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-05-07",
"source": "manifest-variant-identity",
"notes": "F16 activation sibling over the verified Gemma 4 E2B INT4 PLE weight pack. Package-visible runtime evidence is staged at reports/release-claims/gemma-4-e2b-it-q4k-ehf16-af16-int4ple/2026-05-07T20-15-34.710Z.json; Apple Metal capability resolution still fails closed to the AF32 primary when this lane is incompatible.",
"contracts": {
"executionContractOk": true
}
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/gemma-4-E2B-it-ONNX",
"dtype": "q4f16"
}
},
"sourceCheckpointId": "google/gemma-4-E2B-it",
"weightPackId": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple-wp-catalog-v1",
"manifestVariantId": "gemma-4-e2b-it-q4k-ehf16-af16-int4ple-mv-f16-activations-v1",
"artifactCompleteness": "weights-ref",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": true
},
{
"modelId": "translategemma-4b-it-q4k-ehf16-af32",
"family": "translategemma",
"classification": {
"domain": "language",
"tasks": [
"translation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text"
],
"outputs": [
"text"
]
},
"label": "TranslateGemma 4B Instruct (Q4K)",
"description": "TranslateGemma 4B translation model. Q4K weights, F16 embeddings/head, F32 accumulation.",
"modes": [
"translate"
],
"aliases": [
"translategemma-4b",
"translategemma",
"google/translategemma-4b-it",
"translategemma-4b-it-wq4k-ef16-hf16"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "6fc46049882e961a57d1690ba1ffde21677d001a",
"path": "models/translategemma-4b-it-q4k-ehf16-af32"
},
"baseUrl": null,
"sizeBytes": 3167500395,
"recommended": false,
"sortOrder": 15,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "compare",
"tested": "verified"
},
"tested": {
"suite": "debug",
"surface": [
"browser",
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-03-20",
"source": "manual-review",
"notes": "Browser/WebGPU on Apple M3 and Node/WebGPU on AMD RDNA-3 verified through 2026-03-20 with structured TranslateGemma requests. en->fr 'Hello world.' -> 'Bonjour le monde.' The working Q4K path is gemma3-q4k-dequant-f32w-f32a-online; the prior failure remained isolated to the older q4k_dequant F16-weight path.",
"contracts": {
"executionContractOk": true
}
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/translategemma-text-4b-it-ONNX",
"dtype": "q4f16"
}
},
"sourceCheckpointId": "google/translategemma-4b-it",
"weightPackId": "translategemma-4b-it-q4k-ehf16-af32-wp-catalog-v1",
"manifestVariantId": "translategemma-4b-it-q4k-ehf16-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "qwen-3-embedding-0-6b-q4k-ehf16-af32",
"family": "qwen3",
"classification": {
"domain": "language",
"tasks": [
"embedding"
],
"architectureRole": "decoder-embedding-encoder",
"inputs": [
"text"
],
"outputs": [
"pooled-embedding"
]
},
"label": "Qwen3 Embedding 0.6B (Q4K/F32a)",
"description": "Qwen3 0.6B SentenceTransformers embedding target with manifest-owned last-token pooling, prompt inclusion, and L2 normalization. Node/WebGPU verification passes semantic retrieval and pair thresholds; the local compare lane cites the Transformers.js ONNX WebGPU baseline.",
"modes": [
"embedding"
],
"aliases": [
"qwen3-embedding-0.6b",
"qwen3-embedding-0-6b",
"Qwen/Qwen3-Embedding-0.6B"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "049000f49325dca7db2ed2c9de2c8881bd0f4603",
"path": "models/qwen-3-embedding-0-6b-q4k-ehf16-af32"
},
"baseUrl": null,
"quickstart": false,
"demoVisible": false,
"sizeBytes": 1191586416,
"recommended": false,
"sortOrder": 19,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "embedding",
"surface": [
"browser",
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-07-05",
"source": "runtime-verify",
"contracts": {
"executionContractOk": true
},
"notes": "Fresh hosted registry verify receipts pass on node (reports/release-claims/qwen-3-embedding-0-6b-q4k-ehf16-af32/2026-07-05T16-24-38.048Z.node.json) and browser (reports/release-claims/qwen-3-embedding-0-6b-q4k-ehf16-af32/2026-07-05T16-26-35.664Z.browser.json). Semantic retrieval top-1 accuracy 0.90, semantic pair accuracy 0.7143, finite ratio 1.0, 1024-dimensional unit-norm embeddings. Fresh speed evidence: node embeddingMs 144.14, browser embeddingMs 184.70, and local comparable benchmark benchmarks/vendors/results/embedding_compare_qwen-3-embedding-0-6b-q4k-ehf16-af32_20260705T152659.json (Doppler medianEmbeddingMs 67.60; Transformers.js q4f16 medianEmbeddingMs 155.68; correctnessOk true)."
}
},
"verify": {
"workload": "embedding",
"runtimeConfig": {
"inference": {
"prompt": "A small red cube rests on a wooden table beside a blue sphere."
}
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/Qwen3-Embedding-0.6B-ONNX",
"dtype": "q4f16"
}
},
"sourceCheckpointId": "Qwen/Qwen3-Embedding-0.6B",
"weightPackId": "qwen-3-embedding-0-6b-q4k-ehf16-af32-wp-catalog-v1",
"manifestVariantId": "qwen-3-embedding-0-6b-q4k-ehf16-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "qwen-3-reranker-0-6b-q4k-ehf16-af32",
"family": "qwen3",
"classification": {
"domain": "language",
"tasks": [
"reranking"
],
"architectureRole": "decoder-cross-encoder",
"inputs": [
"text-pair"
],
"outputs": [
"relevance-score"
]
},
"label": "Qwen3 Reranker 0.6B (Q4K/F32a)",
"description": "Qwen3 0.6B SentenceTransformers CrossEncoder reranker target with Q4K attention, F16 MLP projections, and manifest-owned true-logit scoring. The hosted registry lane passes semantic rerank verification and has an external Transformers.js ONNX compare profile for fair rerank receipts.",
"modes": [
"rerank"
],
"aliases": [
"qwen3-reranker-0.6b-q4k",
"qwen3-reranker-0-6b-q4k",
"Qwen/Qwen3-Reranker-0.6B:q4k"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "f86fe245b9bbc275cd69af46b1d45d47ea685a55",
"path": "models/qwen-3-reranker-0-6b-q4k-ehf16-af32"
},
"baseUrl": null,
"quickstart": false,
"demoVisible": false,
"sizeBytes": 938321920,
"recommended": false,
"sortOrder": 19.1,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "rerank",
"surface": [
"browser",
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-07-05",
"source": "runtime-verify",
"contracts": {
"executionContractOk": true
},
"notes": "Fresh hosted registry verify receipts pass on node (reports/release-claims/qwen-3-reranker-0-6b-q4k-ehf16-af32/2026-07-05T16-28-17.169Z.node.json) and browser (reports/release-claims/qwen-3-reranker-0-6b-q4k-ehf16-af32/2026-07-05T16-30-58.312Z.browser.json). Semantic pair accuracy 1.0 (5/5) on both receipts, WebGPU query topDocumentIndex 0, rerankMs 1665.34 on node and 1797.80 in browser for the three-document smoke request. Speed evidence: reports/release-claims/qwen-3-reranker-0-6b-q4k-ehf16-af32/2026-07-04T18-43-20.530Z.bench.json (medianRerankMs 1546.27 for the same request). External reranker receipts compare against onnx-community/Qwen3-Reranker-0.6B-ONNX using the q4 ONNX artifact and must not be summarized as a Doppler speed win unless the saved receipt says Doppler leads."
}
},
"verify": {
"workload": "rerank",
"runtimeProfile": "profiles/rerank-stability",
"runtimeConfig": {
"inference": {
"rerank": {
"query": "What API exposes GPU hardware for browser compute workloads?",
"documents": [
"WebGPU provides low-level access to graphics processors for compute workloads.",
"WebSocket provides full-duplex communication over TCP.",
"CSS grid arranges page layout rows and columns."
]
}
}
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/Qwen3-Reranker-0.6B-ONNX",
"dtype": "q4",
"format": "onnx"
}
},
"sourceCheckpointId": "Qwen/Qwen3-Reranker-0.6B",
"weightPackId": "qwen-3-reranker-0-6b-q4k-ehf16-af32-wp-catalog-v2",
"manifestVariantId": "qwen-3-reranker-0-6b-q4k-ehf16-af32-mv-exec-v2",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "qwen-3-reranker-0-6b-f16-af32",
"family": "qwen3",
"classification": {
"domain": "language",
"tasks": [
"reranking"
],
"architectureRole": "decoder-cross-encoder",
"inputs": [
"text-pair"
],
"outputs": [
"relevance-score"
]
},
"label": "Qwen3 Reranker 0.6B (F16/F32a)",
"description": "Qwen3 0.6B SentenceTransformers CrossEncoder reranker target with manifest-owned true/false token scoring. The F16 lane passes the semantic rerank suite and is the supported reranker variant for publication.",
"modes": [
"rerank"
],
"aliases": [
"qwen3-reranker-0.6b",
"qwen3-reranker-0-6b",
"qwen3-reranker-0.6b-f16",
"Qwen/Qwen3-Reranker-0.6B"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "cc1fafff8cda609372b608cd92e487f6a2c32bc8",
"path": "models/qwen-3-reranker-0-6b-f16-af32"
},
"baseUrl": null,
"quickstart": false,
"demoVisible": false,
"sizeBytes": 1191553024,
"recommended": false,
"sortOrder": 19.2,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "rerank",
"surface": [
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-07-04",
"source": "runtime-verify",
"contracts": {
"executionContractOk": true
},
"notes": "Release evidence: reports/release-claims/qwen-3-reranker-0-6b-f16-af32/2026-07-04T02-00-00.000Z.json. Semantic pair accuracy 1.0 (5/5), WebGPU query topDocumentIndex 0, rerankMs 4057.56 for the three-document smoke request. No separate benchmark receipt is promoted for this f16 reranker lane."
}
},
"verify": {
"workload": "rerank",
"runtimeConfig": {
"inference": {
"rerank": {
"query": "What API exposes GPU hardware for browser compute workloads?",
"documents": [
"WebGPU provides low-level access to graphics processors for compute workloads.",
"WebSocket provides full-duplex communication over TCP.",
"CSS grid arranges page layout rows and columns."
]
}
}
}
},
"vendorBenchmark": null,
"sourceCheckpointId": "Qwen/Qwen3-Reranker-0.6B",
"weightPackId": "qwen-3-reranker-0-6b-f16-af32-wp-catalog-v1",
"manifestVariantId": "qwen-3-reranker-0-6b-f16-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "qwen-3-5-0-8b-q4k-ehaf16",
"family": "qwen3",
"classification": {
"domain": "language",
"tasks": [
"generation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text"
],
"outputs": [
"text"
]
},
"label": "Qwen 3.5 0.8B (Q4K)",
"description": "Qwen 3.5 0.8B text model with hybrid linear attention. Q4K weights, F16 embeddings, Q4 tied LM head, F32 accumulation.",
"modes": [
"text"
],
"aliases": [
"qwen3-0.8b",
"qwen-3-5-0-8b",
"Qwen/Qwen3.5-0.8B",
"qwen-3-5-0-8b-wq4k-ef16-hf16-f16"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "80d7716270b6371d541de979eff3370edaf34e13",
"path": "models/qwen-3-5-0-8b-q4k-ehaf16"
},
"baseUrl": null,
"quickstart": true,
"demoVisible": true,
"demoLabel": "Qwen 3.5 0.8B",
"demoRole": "Compact general text",
"demoSortOrder": 20,
"sizeBytes": 1155348912,
"recommended": false,
"sortOrder": 20,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "inference",
"surface": [
"browser",
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-06-15",
"source": "local-node-webgpu-verify",
"notes": "Browser/Node release evidence is package-visible at reports/release-claims/qwen-3-5-0-8b-q4k-ehaf16/2026-05-10T02-22-04.891Z.json. The hosted clocksmith/rdrr artifact is approved for quickstart resolution with explicit stop-token metadata.",
"contracts": {
"executionContractOk": true
}
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/Qwen3.5-0.8B-ONNX",
"dtype": "q4f16"
}
},
"verify": {
"workload": "inference",
"runtimeConfig": {
"inference": {
"prompt": "The color of the sky is",
"generation": {
"maxTokens": 8
},
"sampling": {
"temperature": 0,
"topK": 1,
"topP": 1,
"repetitionPenalty": 1,
"greedyThreshold": 0
}
}
}
},
"sourceCheckpointId": "Qwen/Qwen3.5-0.8B",
"weightPackId": "qwen-3-5-0-8b-q4k-ehaf16-wp-catalog-v1",
"manifestVariantId": "qwen-3-5-0-8b-q4k-ehaf16-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "qwen-3-5-2b-q4k-ehaf16",
"family": "qwen3",
"classification": {
"domain": "language",
"tasks": [
"generation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text"
],
"outputs": [
"text"
]
},
"label": "Qwen 3.5 2B (Q4K)",
"description": "Qwen 3.5 2B text model with hybrid linear attention. Q4K weights, F16 embeddings, Q4 tied LM head, F32 accumulation.",
"modes": [
"text"
],
"aliases": [
"qwen3-2b",
"qwen-3-5-2b",
"Qwen/Qwen3.5-2B",
"qwen-3-5-2b-wq4k-ef16-hf16-f16"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "977d145bf2478a7fb542e6aca65030585620ca60",
"path": "models/qwen-3-5-2b-q4k-ehaf16"
},
"baseUrl": null,
"quickstart": true,
"demoVisible": true,
"demoLabel": "Qwen 3.5 2B",
"demoRole": "Higher-capacity text",
"demoSortOrder": 30,
"sizeBytes": 2496068973,
"recommended": false,
"sortOrder": 21,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "inference",
"surface": [
"browser",
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-06-15",
"source": "local-node-webgpu-verify",
"notes": "Browser/Node release evidence is package-visible at reports/release-claims/qwen-3-5-2b-q4k-ehaf16/2026-05-03T02-33-21.397Z.json. The execution-v1 lane keeps the fixed fused-Q4 decode path and q4 LM-head/prefill kernels; the hosted clocksmith/rdrr artifact is approved for quickstart resolution with explicit stop-token metadata.",
"contracts": {
"executionContractOk": true
}
}
},
"vendorBenchmark": {
"transformersjs": {
"repoId": "onnx-community/Qwen3.5-2B-ONNX",
"dtype": "q4f16"
}
},
"sourceCheckpointId": "Qwen/Qwen3.5-2B",
"weightPackId": "qwen-3-5-2b-q4k-ehaf16-wp-catalog-v1",
"manifestVariantId": "qwen-3-5-2b-q4k-ehaf16-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "qwen-3-6-27b-q4k-ehaf16",
"family": "qwen3",
"classification": {
"domain": "language",
"tasks": [
"generation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text"
],
"outputs": [
"text"
]
},
"label": "Qwen 3.6 27B (Q4K)",
"description": "Qwen 3.6 27B text model with hybrid linear attention. Q4K weights, F16 embeddings, F32 accumulation.",
"modes": [
"text"
],
"aliases": [
"qwen3.6-27b",
"qwen-3-6-27b",
"Qwen/Qwen3.6-27B"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "b402f6f27837857d51636da5f78c12bcd47e2a03",
"path": "models/qwen-3-6-27b-q4k-ehaf16"
},
"baseUrl": null,
"quickstart": false,
"demoVisible": false,
"demoPreferredVariantId": "qwen-3-6-27b-q4k-eaf16",
"demoWarningBadges": [
"15.8 GB / High RAM"
],
"demoWarningText": "Large download. Use a high-RAM machine.",
"sizeBytes": 16994352128,
"recommended": false,
"sortOrder": 22,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "inference",
"surface": [
"browser"
],
"result": "pass",
"lastVerifiedAt": "2026-04-28",
"source": "browser-webgpu-smoke",
"notes": "Committed program-bundle reference reports/program-bundles/qwen-3-6-27b-q4k-ehaf16/2026-04-28T01-19-10.497Z.reference.json captured a deterministic reference transcript and executionContractOk=true. Hosted artifact is enabled for web demo download; quickstart remains disabled.",
"contracts": {
"executionContractOk": true
}
}
},
"vendorBenchmark": null,
"sourceCheckpointId": "Qwen/Qwen3.6-27B",
"weightPackId": "qwen-3-6-27b-q4k-ehaf16-wp-catalog-v1",
"manifestVariantId": "qwen-3-6-27b-q4k-ehaf16-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "qwen-3-6-27b-q4k-eaf16",
"family": "qwen3",
"classification": {
"domain": "language",
"tasks": [
"generation"
],
"architectureRole": "autoregressive-decoder",
"inputs": [
"text"
],
"outputs": [
"text"
]
},
"label": "Qwen 3.6 27B (Q4K/F16a)",
"description": "Qwen 3.6 27B text-only Doppler/WebGPU f16 lane. Shares the existing Q4K/F16 embedding weight pack; Doe/Cerebras f16 CSL lowering is intentionally unsupported.",
"modes": [
"text"
],
"aliases": [
"qwen3.6-27b-f16",
"qwen-3-6-27b-af16"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"revision": "3dee21b3b12d65ac7fef9b24cbf759cacc953a67",
"path": "models/qwen-3-6-27b-q4k-eaf16"
},
"baseUrl": null,
"quickstart": false,
"demoVisible": false,
"sizeBytes": 16994352128,
"recommended": false,
"sortOrder": 23,
"lifecycle": {
"availability": {
"curated": false,
"local": true,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "inference",
"surface": [
"browser",
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-04-29",
"source": "browser-node-webgpu-smoke",
"notes": "Committed program-bundle reference reports/program-bundles/qwen-3-6-27b-q4k-eaf16/capture.node.reference.json produces coherent sky prompt output with f16 compute.",
"contracts": {
"executionContractOk": true
}
}
},
"vendorBenchmark": null,
"sourceCheckpointId": "Qwen/Qwen3.6-27B",
"weightPackId": "qwen-3-6-27b-q4k-ehaf16-wp-catalog-v1",
"manifestVariantId": "qwen-3-6-27b-q4k-eaf16-mv-exec-v1",
"artifactCompleteness": "weights-ref",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": true
},
{
"modelId": "amplify-120m-f16-af32",
"family": "amplify",
"classification": {
"domain": "protein",
"tasks": [
"embedding",
"masked-token-prediction"
],
"architectureRole": "bidirectional-encoder",
"inputs": [
"protein-sequence"
],
"outputs": [
"token-embedding",
"pooled-embedding",
"token-logits"
]
},
"label": "AMPLIFY 120M Protein Encoder (F16/F32a)",
"description": "AMPLIFY 120M protein-sequence encoder with F16 weights and F32 activations.",
"modes": [
"embedding"
],
"aliases": [
"amplify-120m",
"chandar-lab/AMPLIFY_120M"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"path": "models/amplify-120m-f16-af32",
"revision": "990de206946a0972cc92e9dbcc9726399d257cac"
},
"baseUrl": null,
"quickstart": false,
"sizeBytes": 236647491,
"recommended": false,
"sortOrder": 36,
"lifecycle": {
"availability": {
"curated": false,
"local": false,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "sequence-embedding",
"surface": [
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-07-19",
"source": "runtime-verify",
"notes": "Hosted Node/WebGPU sequence parity and synthetic q_proj LoRA lifecycle passed at docs/status/amplify-120m-sequence-hosted-webgpu-lora-qualification-2026-07-19.json. This is not a downstream biological-task quality claim.",
"contracts": {
"executionContractOk": true
}
}
},
"sourceCheckpointId": "chandar-lab/AMPLIFY_120M@d918a9e8c64dc43f6ec4996022c3deba14d66470",
"weightPackId": "amplify-120m-f16-af32-wp-v1",
"manifestVariantId": "amplify-120m-f16-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "esm2-t12-35m-ur50d-f32-af32",
"family": "esm",
"classification": {
"domain": "protein",
"tasks": [
"embedding"
],
"architectureRole": "bidirectional-encoder",
"inputs": [
"protein-sequence"
],
"outputs": [
"token-embedding",
"pooled-embedding"
]
},
"label": "ESM-2 35M Protein Encoder (F32)",
"description": "ESM-2 35M protein-sequence encoder with F32 weights and activations.",
"modes": [
"embedding"
],
"aliases": [
"esm2-35m",
"facebook/esm2_t12_35M_UR50D"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"path": "models/esm2-t12-35m-ur50d-f32-af32",
"revision": "feea324bae6c06c9d129ce39e188a75478ccdbbc"
},
"baseUrl": null,
"quickstart": false,
"sizeBytes": 133159807,
"recommended": false,
"sortOrder": 37,
"lifecycle": {
"availability": {
"curated": false,
"local": false,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "sequence-embedding",
"surface": [
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-07-19",
"source": "runtime-verify",
"notes": "Hosted Node/WebGPU sequence parity and synthetic q_proj LoRA lifecycle passed at docs/status/esm2-t12-35m-sequence-hosted-webgpu-lora-qualification-2026-07-19.json. The initial lane exposes embeddings, not masked-token logits or downstream biological quality.",
"contracts": {
"executionContractOk": true
}
}
},
"sourceCheckpointId": "facebook/esm2_t12_35M_UR50D@6fbf070e65b0b7291e7bbcd451118c216cff79d8",
"weightPackId": "esm2-t12-35m-ur50d-f32-af32-wp-v1",
"manifestVariantId": "esm2-t12-35m-ur50d-f32-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "esmc-300m-f32-af32",
"family": "esmc",
"classification": {
"domain": "protein",
"tasks": [
"embedding"
],
"architectureRole": "bidirectional-encoder",
"inputs": [
"protein-sequence"
],
"outputs": [
"token-embedding",
"pooled-embedding"
]
},
"label": "ESMC 300M Protein Encoder (F32)",
"description": "ESMC 300M protein-sequence encoder with F32 weights and activations.",
"modes": [
"embedding"
],
"aliases": [
"esmc-300m",
"biohub/ESMC-300M"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"path": "models/esmc-300m-f32-af32",
"revision": "8354795dc544798ca7a4419bf00f844648f4b6f0"
},
"baseUrl": null,
"quickstart": false,
"sizeBytes": 1328207519,
"recommended": false,
"sortOrder": 38,
"lifecycle": {
"availability": {
"curated": false,
"local": false,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "sequence-embedding",
"surface": [
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-07-19",
"source": "runtime-verify",
"notes": "Hosted Node/WebGPU sequence parity and synthetic q_proj LoRA lifecycle passed at docs/status/esmc-300m-sequence-hosted-webgpu-lora-qualification-2026-07-19.json. The initial lane exposes embeddings, not masked-token logits or downstream biological quality.",
"contracts": {
"executionContractOk": true
}
}
},
"sourceCheckpointId": "biohub/ESMC-300M@a59b831785f907e96e6a246b1d142bfb76df31ee",
"weightPackId": "esmc-300m-f32-af32-wp-v1",
"manifestVariantId": "esmc-300m-f32-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
},
{
"modelId": "nucleotide-transformer-v2-50m-f32-af32",
"family": "nucleotide-transformer",
"classification": {
"domain": "nucleotide",
"tasks": [
"embedding"
],
"architectureRole": "bidirectional-encoder",
"inputs": [
"dna-sequence"
],
"outputs": [
"token-embedding",
"pooled-embedding"
]
},
"label": "Nucleotide Transformer v2 50M DNA Encoder (F32)",
"description": "Nucleotide Transformer v2 50M DNA-sequence encoder with F32 weights and activations.",
"modes": [
"embedding"
],
"aliases": [
"nucleotide-transformer-v2-50m",
"InstaDeepAI/nucleotide-transformer-v2-50m-multi-species"
],
"artifact": {
"format": "rdrr"
},
"hf": {
"repoId": "clocksmith/rdrr",
"path": "models/nucleotide-transformer-v2-50m-f32-af32",
"revision": "a5703ef6971cc883438a393d73f2f54dccf94976"
},
"baseUrl": null,
"quickstart": false,
"sizeBytes": 210074179,
"recommended": false,
"sortOrder": 39,
"lifecycle": {
"availability": {
"curated": false,
"local": false,
"hf": true
},
"status": {
"runtime": "active",
"conversion": "ready",
"demo": "none",
"tested": "verified"
},
"tested": {
"suite": "sequence-embedding",
"surface": [
"node"
],
"result": "pass",
"lastVerifiedAt": "2026-07-19",
"source": "runtime-verify",
"notes": "Hosted Node/WebGPU sequence parity and synthetic q_proj LoRA lifecycle passed at docs/status/nucleotide-transformer-v2-50m-sequence-hosted-webgpu-lora-qualification-2026-07-19.json. The initial lane exposes embeddings, not masked-token logits or downstream biological quality; upstream use is CC-BY-NC-SA-4.0.",
"contracts": {
"executionContractOk": true
}
}
},
"sourceCheckpointId": "InstaDeepAI/nucleotide-transformer-v2-50m-multi-species@81b29e5786726d891dbf929404ef20adca5b36f1",
"weightPackId": "nucleotide-transformer-v2-50m-f32-af32-wp-v1",
"manifestVariantId": "nucleotide-transformer-v2-50m-f32-af32-mv-exec-v1",
"artifactCompleteness": "complete",
"runtimePromotionState": "manifest-owned",
"weightsRefAllowed": false
}
]
}