{
  "version": 1,
  "einheiten": "Decimal: 1 GB = 1,000,000,000 bytes; 1 TB = 1,000 GB.",
  "hardware_heute": "NVIDIA GeForce RTX 4080 (16 GB VRAM)",
  "hardware_plan": "Two Mac Studio machines, 512 GB each; not delivered.",
  "rest_hinweis": "Budget minus model files leaves an upper bound for context, runtime buffers and the operating system; no context length or runtime compatibility is promised.",
  "slug": "gemma-4-12b-qat-q4-0",
  "anzeige": "Gemma 4 12B QAT",
  "hf_repo": null,
  "variante": "Q4_0",
  "groesse": {
    "bytes": null,
    "gb_gerundet": "7.15",
    "dateien": null,
    "quelle": "MESSUNG_2026-10-13.json",
    "abgerufen": "2026-10-26",
    "hinweis": "Local model-file size recorded on 13 October 2026, rounded to 0.01 GB; exact bytes and Hugging Face repository were not recorded."
  },
  "budgets": [
    {
      "id": "24gb",
      "anzeige": "24 GB",
      "plan": false,
      "hinweis": "Comparison budget; the measured RTX 4080 has 16 GB VRAM.",
      "passt": true,
      "rest_bytes": "16850000000",
      "gerundete_grundlage": true
    },
    {
      "id": "512gb",
      "anzeige": "512 GB",
      "plan": true,
      "hinweis": "One planned Mac Studio; not delivered.",
      "passt": true,
      "rest_bytes": "504850000000",
      "gerundete_grundlage": true
    },
    {
      "id": "1tb",
      "anzeige": "1 TB",
      "plan": true,
      "hinweis": "Conservative comparison budget across two planned 512 GB Macs; not delivered. Not one shared memory pool.",
      "passt": true,
      "rest_bytes": "992850000000",
      "gerundete_grundlage": true
    }
  ],
  "messungen": [
    {
      "id": "MESSUNG_2026-10-13",
      "quelle": "MESSUNG_2026-10-13.json",
      "zeit": "2026-10-13T13:30:09",
      "zeitzone": null,
      "hardware": "NVIDIA GeForce RTX 4080",
      "vram_gesamt_mib": 16376,
      "software": "LM Studio",
      "software_version": null,
      "kontext_tokens": null,
      "methode": {
        "temperatur": null,
        "max_tokens": 300,
        "wiederholungen": 2,
        "aggregation": "recorded medians",
        "prompts": "Fresh identifier and topic per request; no prompt cache.",
        "exakte_prompts": null
      },
      "laeufe": [
        {
          "id": "kurz",
          "tok_s": 74,
          "ttft_s": 0.14,
          "prompt_tokens": 37.5,
          "watt": 161.2,
          "tok_s_einzeln": [
            73.4,
            74.7
          ]
        },
        {
          "id": "mittel",
          "tok_s": 72.8,
          "ttft_s": 0.42,
          "prompt_tokens": 870,
          "watt": 178,
          "tok_s_einzeln": [
            73,
            72.7
          ]
        },
        {
          "id": "lang",
          "tok_s": 66.1,
          "ttft_s": 1.01,
          "prompt_tokens": 3449.5,
          "watt": 185.2,
          "tok_s_einzeln": [
            66,
            66.2
          ]
        }
      ],
      "eur_pro_mio_tokens": 0.21,
      "strompreis_eur_kwh": 0.35,
      "stromumfang": "GPU only",
      "vram_belegt_gb": 8.9,
      "rohurl": "/lab/gemma-4-12b-qat-q4-0/data.json"
    },
    {
      "id": "MESSUNG_2026-10-27",
      "quelle": "MESSUNG_2026-10-27.json",
      "zeit": "2026-10-27T13:30:02+01:00",
      "zeitzone": "recorded in timestamp",
      "hardware": "NVIDIA GeForce RTX 4080",
      "vram_gesamt_mib": 16376,
      "software": "LM Studio",
      "software_version": "2.61.0",
      "kontext_tokens": 16384,
      "methode": {
        "temperatur": null,
        "max_tokens": 1500,
        "wiederholungen": 2,
        "aggregation": "recorded medians",
        "prompts": "Fresh identifier and topic per request; no prompt cache.",
        "exakte_prompts": null
      },
      "laeufe": [
        {
          "id": "kurz",
          "tok_s": 73.9,
          "ttft_s": 0.13,
          "prompt_tokens": 41,
          "watt": 235.8,
          "tok_s_einzeln": [
            73.9,
            73.8
          ],
          "ttft_wand_s": 2.12,
          "ttft_wand_methode": "hergeleitet: dauer_s - gen_s (stream:false)",
          "ttft_engine_s": 0.13,
          "kosten_eur_1m_tokens": "0.3102165087956698240866035183",
          "strom_leistungsfeld": "power.draw.instant",
          "tokenbasis": "alle erzeugten Tokens einschliesslich Reasoning (completion_tokens)",
          "antwort_tokens": 1136.5,
          "reasoning_tokens": 577,
          "sichtbare_tokens": 559.5
        },
        {
          "id": "mittel",
          "tok_s": 72,
          "ttft_s": 0.47,
          "prompt_tokens": 863.5,
          "watt": 235,
          "tok_s_einzeln": [
            72.1,
            71.9
          ],
          "ttft_wand_s": 2.23,
          "ttft_wand_methode": "hergeleitet: dauer_s - gen_s (stream:false)",
          "ttft_engine_s": 0.47,
          "kosten_eur_1m_tokens": "0.3173225308641975308641975309",
          "strom_leistungsfeld": "power.draw.instant",
          "tokenbasis": "alle erzeugten Tokens einschliesslich Reasoning (completion_tokens)",
          "antwort_tokens": 827.5,
          "reasoning_tokens": 654,
          "sichtbare_tokens": 173.5
        },
        {
          "id": "lang",
          "tok_s": 65.3,
          "ttft_s": 1.02,
          "prompt_tokens": 3460,
          "watt": 239.6,
          "tok_s_einzeln": [
            65.3,
            65.2
          ],
          "ttft_wand_s": 2.35,
          "ttft_wand_methode": "hergeleitet: dauer_s - gen_s (stream:false)",
          "ttft_engine_s": 1.02,
          "kosten_eur_1m_tokens": "0.3567296239578015994555045091",
          "strom_leistungsfeld": "power.draw.instant",
          "tokenbasis": "alle erzeugten Tokens einschliesslich Reasoning (completion_tokens)",
          "antwort_tokens": 927.5,
          "reasoning_tokens": 741.5,
          "sichtbare_tokens": 186
        }
      ],
      "eur_pro_mio_tokens": 0.31,
      "strompreis_eur_kwh": 0.35,
      "stromumfang": "GPU only",
      "vram_belegt_gb": 8.9,
      "rohurl": "/lab/gemma-4-12b-qat-q4-0/data.json",
      "software_version_quelle": "engine",
      "engine": "llama.cpp-win-x86_64-nvidia-cuda12-avx2",
      "engine_version": "2.61.0",
      "strom_leistungsfeld": "power.draw.instant",
      "tokenmethode": "native stats.tokens_per_second; alle erzeugten Tokens einschliesslich Reasoning (completion_tokens)"
    }
  ]
}