{
  "$comment": "Verified LlamaBox device/model benchmark records. Mirror of the table on /tested-devices.html. EVERY entry must come from a real measured run on real hardware. Do not interpolate, estimate, or copy the indicative ranges published on the homepage performance panel. An empty array is honest; a fabricated row is not.",
  "schemaVersion": 1,
  "lastUpdated": "2026-09-16",
  "methodologyUrl": "https://llamabox-ai.vercel.app/tested-devices.html",
  "fields": {
    "device": "Marketing name and model number, e.g. 'Pixel 6a (GX7AS)'",
    "soc": "System-on-chip identifier, e.g. 'Google Tensor'",
    "androidVersion": "Android release and API level, e.g. '14 (API 34)'",
    "ramGb": "Total physical RAM in GB (number)",
    "model": "Model name and parameter count, e.g. 'Qwen2.5 0.5B Instruct'",
    "quantization": "GGUF quantization, e.g. 'Q4_K_M'",
    "modelSha256": "SHA-256 of the exact GGUF file tested",
    "contextLength": "Context window used for the run (number)",
    "threads": "Inference thread count (number)",
    "loadTimeSec": "Seconds from load start to model ready (number)",
    "timeToFirstTokenSec": "Latency to first generated token (number)",
    "promptTokPerSec": "Prompt-processing throughput (number)",
    "generationTokPerSec": "Generation throughput (number)",
    "peakMemoryMb": "Peak resident memory observed in MB (number)",
    "thermalNotes": "Throttling or heat observations (string, may be empty)",
    "result": "One of: ok | slow | unstable | failed-oom | failed-unsupported",
    "appVersion": "LlamaBox versionName under test, e.g. '1.0.7'",
    "testDate": "ISO 8601 date of the run, e.g. '2026-08-26'"
  },
  "resultValues": [
    "ok",
    "slow",
    "unstable",
    "failed-oom",
    "failed-unsupported"
  ],
  "exampleStructure": {
    "$comment": "SHAPE REFERENCE ONLY — this is not a measurement and must never be rendered as published data. Delete or ignore when adding real records.",
    "device": "<device name and model number>",
    "soc": "<system-on-chip>",
    "androidVersion": "<release> (API <level>)",
    "ramGb": 0,
    "model": "<model name and parameter count>",
    "quantization": "<e.g. Q4_K_M>",
    "modelSha256": "<64-character hex digest>",
    "contextLength": 0,
    "threads": 0,
    "loadTimeSec": 0,
    "timeToFirstTokenSec": 0,
    "promptTokPerSec": 0,
    "generationTokPerSec": 0,
    "peakMemoryMb": 0,
    "thermalNotes": "",
    "result": "ok",
    "appVersion": "<version>",
    "testDate": "<YYYY-MM-DD>"
  },
  "records": [],
  "caveats": [
    "Results vary by CPU, memory bandwidth, thermal conditions, context length and model architecture.",
    "A model fitting in storage does not guarantee it will fit in usable RAM.",
    "Q4_K_M is a recommendation, not universal compatibility.",
    "Vision requires a compatible multimodal model and a matching projector.",
    "Android may reclaim memory and interrupt generation under pressure, independently of LlamaBox.",
    "Failed runs are published alongside successful ones; they are the most useful rows in the table."
  ]
}