{ "$comment": "GENERATED for curated-public rows by scripts/seed-curated-benchmarks.ts. Community submissions are merged in by scripts/export-measured-benchmarks.ts and carry provenance \"community-submitted\".", "schemaVersion": "2.0", "schemaDescription": "Each entry is one benchmark run of one model on one accelerator at one quantization and context length. tokS is decode throughput. Fields added in v2 (promptTokS, vramUsedGb, measuredWatts + wattsMethod, offloadState, nGpuLayers, kvCacheType, batchSize, reviewedAt) are optional and null where no figure exists. A record carrying measuredWatts always carries wattsMethod; a community-submitted record always carries reviewedAt.", "generatedAt": "2026-09-08T15:39:01.768Z", "license": "CC BY 4.0", "attribution": "LLM Configurator — https://llmconfigurator.com/en/benchmarks?utm_source=dataset&utm_medium=referral&utm_campaign=measured_benchmarks_export", "provenanceNote": "Rows marked curated-public are compiled from third-party published benchmarks and are credited to their publisher with a source URL; they are not submissions to this site. Rows marked community-submitted come from readers and carry a sample count.", "entries": [ { "id": "hc-4090-qwen3-30b-a3b-4k", "gpuId": "rtx-4090", "variantId": "qwen3-30b-a3b", "quant": "Q4_K_XL", "runtime": "llama.cpp llama-bench", "contextLength": 4096, "os": "unspecified", "tokS": 195.78, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/", "modelIdentity": "confirmed", "note": "Source states the 30B MoE is 16.47 GB at Q4_K_XL, which matches Qwen3-30B-A3B and fixes the file size directly rather than deriving it from a bits-per-weight table.", "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "hc-4090-qwen3-30b-a3b-57k", "gpuId": "rtx-4090", "variantId": "qwen3-30b-a3b", "quant": "Q4_K_XL", "runtime": "llama.cpp llama-bench", "contextLength": 57344, "os": "unspecified", "tokS": 74.64, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/", "modelIdentity": "confirmed", "note": "57K is the largest context the source reports fitting on a 24 GB card for this model.", "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "hc-4090-qwen3-8b-4k", "gpuId": "rtx-4090", "variantId": "qwen3-8b", "quant": "Q4_K_M", "runtime": "llama.cpp llama-bench", "contextLength": 4096, "os": "unspecified", "tokS": 141.3, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/", "modelIdentity": "confirmed", "note": "Named Qwen3-8B Q4_K context-scaling series, five points from 4K to 128K.", "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "hc-4090-qwen3-8b-16k", "gpuId": "rtx-4090", "variantId": "qwen3-8b", "quant": "Q4_K_M", "runtime": "llama.cpp llama-bench", "contextLength": 16384, "os": "unspecified", "tokS": 108, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/", "modelIdentity": "confirmed", "note": null, "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "hc-4090-qwen3-8b-32k", "gpuId": "rtx-4090", "variantId": "qwen3-8b", "quant": "Q4_K_M", "runtime": "llama.cpp llama-bench", "contextLength": 32768, "os": "unspecified", "tokS": 82.3, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/", "modelIdentity": "confirmed", "note": null, "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "hc-4090-qwen3-8b-64k", "gpuId": "rtx-4090", "variantId": "qwen3-8b", "quant": "Q4_K_M", "runtime": "llama.cpp llama-bench", "contextLength": 65536, "os": "unspecified", "tokS": 56.1, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/", "modelIdentity": "confirmed", "note": null, "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "hc-4090-qwen3-8b-128k", "gpuId": "rtx-4090", "variantId": "qwen3-8b", "quant": "Q4_K_M", "runtime": "llama.cpp llama-bench", "contextLength": 131072, "os": "unspecified", "tokS": 33.8, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/gpu-llm-benchmarks/rtx-4090/", "modelIdentity": "confirmed", "note": null, "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "hc-4090-8b-q4kxl-4k", "gpuId": "rtx-4090", "variantId": "llama-3.1-8b", "quant": "Q4_K_XL", "runtime": "llama.cpp llama-bench", "contextLength": 4096, "os": "unspecified", "tokS": 131, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/", "modelIdentity": "inferred", "note": "Source reports this row as '8B' without naming the checkpoint. Attributed to Llama 3.1 8B; its 32-layer/8-KV-head geometry differs slightly from Qwen3-8B's 36 layers, so the identity is marked inferred.", "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "hc-4090-8b-q4kxl-65k", "gpuId": "rtx-4090", "variantId": "llama-3.1-8b", "quant": "Q4_K_XL", "runtime": "llama.cpp llama-bench", "contextLength": 65536, "os": "unspecified", "tokS": 53.07, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/", "modelIdentity": "inferred", "note": "Pairs with the 4K row; the 59% drop between them is the headline context-scaling figure.", "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "hc-4090-14b-q4kxl-4k", "gpuId": "rtx-4090", "variantId": "qwen3-14b", "quant": "Q4_K_XL", "runtime": "llama.cpp llama-bench", "contextLength": 4096, "os": "unspecified", "tokS": 82.82, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/", "modelIdentity": "inferred", "note": "Source reports '14B' only. Attributed to Qwen3-14B because the publisher's confirmed rows on the same page (8B, 30B-A3B) are Qwen3.", "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "hc-4090-14b-q4kxl-65k", "gpuId": "rtx-4090", "variantId": "qwen3-14b", "quant": "Q4_K_XL", "runtime": "llama.cpp llama-bench", "contextLength": 65536, "os": "unspecified", "tokS": 38.73, "provenance": "curated-public", "publisher": "Hardware Corner", "sourceUrl": "https://www.hardware-corner.net/rtx-4090-llm-benchmarks/", "modelIdentity": "inferred", "note": null, "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "mustafa-4090-7b", "gpuId": "rtx-4090", "variantId": "qwen-2.5-7b", "quant": "Q4_K_M", "runtime": "llama.cpp llama-bench", "contextLength": 4096, "os": "unspecified", "tokS": 135, "provenance": "curated-public", "publisher": "Mustafa.net", "sourceUrl": "https://mustafa.net/llm-tokens-per-second-benchmarks/", "modelIdentity": "inferred", "note": "Source reports a generic '7B Q4_K_M, single batch, warm model' row with no checkpoint and no stated context. Qwen 2.5 7B is used as a standard-geometry 7B stand-in and 4K assumed.", "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "mustafa-3090-7b", "gpuId": "rtx-3090", "variantId": "qwen-2.5-7b", "quant": "Q4_K_M", "runtime": "llama.cpp llama-bench", "contextLength": 4096, "os": "unspecified", "tokS": 95, "provenance": "curated-public", "publisher": "Mustafa.net", "sourceUrl": "https://mustafa.net/llm-tokens-per-second-benchmarks/", "modelIdentity": "inferred", "note": "THE ONLY AMPERE ANCHOR IN THE SET. eta for nvidia-ampere rests entirely on this one inferred-identity point; treat the Ampere constant as materially weaker than the Ada one.", "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null }, { "id": "mustafa-m3max-7b", "gpuId": "apple-m3-max", "variantId": "qwen-2.5-7b", "quant": "Q4_K_M", "runtime": "llama.cpp llama-bench", "contextLength": 4096, "os": "unspecified", "tokS": 40, "provenance": "curated-public", "publisher": "Mustafa.net", "sourceUrl": "https://mustafa.net/llm-tokens-per-second-benchmarks/", "modelIdentity": "inferred", "note": "THE ONLY APPLE ANCHOR IN THE SET, and the only Metal row. With one point, eta and t_fixed are not jointly identifiable, so the fit solves eta only and leaves the Metal t_fixed at its seed value, reported unfitted.", "date": "2026-08-17", "offloadState": "fully-resident", "reviewedAt": null, "promptTokS": null, "vramUsedGb": null, "measuredWatts": null, "wattsMethod": null, "measuredNoiseDba": null, "measuredDeltaC": null, "nGpuLayers": null, "kvCacheType": null, "batchSize": null, "submittedBy": null } ] }