{
  "title": "Gemma 4 12B IT QAT Q4_0 on RTX 5090: long-context rerun measured by llm-speed",
  "test_date": "2026-10-22",
  "measured_by": "llm-speed (@llmspeed, https://x.com/llmspeed)",
  "citation": "llm-speed (2026-10-22), Gemma 4 12B IT QAT Q4_0 on RTX 5090: long-context rerun. https://llm-speed.com/m/gemma-4-12b-it-qat#long-context",
  "snapshot_note": "Dated snapshot of nine measurements from three runs measured by llm-speed on an RTX 5090 (listed as confirmed in run_sources). It reruns the September 6 community-submitted runs (gemma-4-5090-context-2026-09-06.json) with the same llama.cpp commit, model revision and file, context and workload. Estimated ctx targets are not actual prompt token counts.",
  "configuration": {
    "hardware": "RTX 5090 (32GB), Linux",
    "driver": "NVIDIA 595.91.07",
    "runtime": "llama.cpp CUDA",
    "runtime_commit": "9725a313be0528214c4a02fed906ddaf7b3f712e",
    "runtime_build": "b8931 (reported as 8931 (9725a313b))",
    "client_version": "llm-speed 0.0.9 (PyPI), --strict-anon",
    "suite": "suite-v1",
    "workload": "long-context-decay",
    "model_repository": "google/gemma-4-12B-it-qat-q4_0-gguf",
    "model_revision": "29d097773436b69ff9feafd636ab4cf873786537",
    "model_file": "gemma-4-12b-it-qat-q4_0.gguf",
    "model_sha256": "93567e57a8fe10b23569b9d9ec38cd005deedf71e29477c421a4b83f418a538b",
    "quantization": "Q4_0",
    "configured_context_tokens": 131072,
    "output_cap_tokens": 256,
    "parallel": 1,
    "gpu_layers": "49/49",
    "prompt_cache": false,
    "ram_cache_mib": 0,
    "thinking": "off",
    "temperature": 0,
    "seed": 42,
    "server_restarted_per_repetition": true,
    "peak_device_vram_mib": 10288
  },
  "limits": [
    "Three sequential repetitions on one configuration; observed ranges are not confidence intervals.",
    "Not a full 128K-input or maximum-context test.",
    "No evaluation of answer quality or multi-user throughput.",
    "Board power and clock limits were not recorded. One llama-server process per repetition.",
    "The GGUF file was renamed per repetition after its sha256 was checked; all three use the same bytes.",
    "Prompt targets use a four-characters-per-token estimate; prompt_tokens_actual is runtime-reported usage."
  ],
  "compared_with": {
    "snapshot": "https://llm-speed.com/data/gemma-4-5090-context-2026-09-06.json",
    "source": "community-submitted, not measured by llm-speed",
    "medians": [
      {
        "actual_prompt_tokens": 23872,
        "decode_tps_change_pct": -4.5,
        "ttft_change_pct": 20.8
      },
      {
        "actual_prompt_tokens": 47652,
        "decode_tps_change_pct": -4.4,
        "ttft_change_pct": 10.1
      },
      {
        "actual_prompt_tokens": 95212,
        "decode_tps_change_pct": 1.4,
        "ttft_change_pct": 8.2
      }
    ],
    "note": "Different machine, operating system, driver and client; the most likely cause of the differences."
  },
  "summary": [
    {
      "target": 32000,
      "actual_prompt_tokens": 23872,
      "ttft_ms": {
        "median": 4052.181963980198,
        "min": 4001.4562849998474,
        "max": 4058.8335899710655
      },
      "decode_tps": {
        "median": 130.15110253900494,
        "min": 129.81955925197605,
        "max": 130.55676502601895
      },
      "prefill_tps": {
        "median": 5891.147093639415,
        "min": 5881.492667988435,
        "max": 5965.828013538053
      },
      "output_tokens": {
        "median": 98,
        "min": 98,
        "max": 98
      }
    },
    {
      "target": 64000,
      "actual_prompt_tokens": 47652,
      "ttft_ms": {
        "median": 8697.730915009975,
        "min": 8673.536598950624,
        "max": 8731.327500998974
      },
      "decode_tps": {
        "median": 127.12279138920348,
        "min": 127.02932424243787,
        "max": 127.19318907621664
      },
      "prefill_tps": {
        "median": 5478.670295233587,
        "min": 5457.589352197361,
        "max": 5493.952721173186
      },
      "output_tokens": {
        "median": 100,
        "min": 100,
        "max": 100
      }
    },
    {
      "target": 128000,
      "actual_prompt_tokens": 95212,
      "ttft_ms": {
        "median": 21539.667991012335,
        "min": 21501.898664981127,
        "max": 21565.837944954634
      },
      "decode_tps": {
        "median": 117.56070820542296,
        "min": 117.18462584586108,
        "max": 117.58587105022652
      },
      "prefill_tps": {
        "median": 4420.309544219914,
        "min": 4414.945537614735,
        "max": 4428.074073061564
      },
      "output_tokens": {
        "median": 96,
        "min": 96,
        "max": 96
      }
    }
  ],
  "measurements": [
    {
      "repetition": 1,
      "source_url": "https://llm-speed.com/r/r_9suyze4n7lx",
      "ctx": 32000,
      "prompt_tokens_actual": 23872,
      "output_tokens": 98,
      "prefill_tps": 5881.492667988435,
      "decode_tps": 130.55676502601895,
      "ttft_ms": 4058.8335899710655,
      "wall_ms": 4803.947223991156,
      "p50_ms": 7.499778002500534,
      "p95_ms": 7.768293261528015,
      "skipped": false,
      "error": null
    },
    {
      "repetition": 1,
      "source_url": "https://llm-speed.com/r/r_9suyze4n7lx",
      "ctx": 64000,
      "prompt_tokens_actual": 47652,
      "output_tokens": 100,
      "prefill_tps": 5493.952721173186,
      "decode_tps": 127.19318907621664,
      "ttft_ms": 8673.536598950624,
      "wall_ms": 9455.394015997648,
      "p50_ms": 7.803710505366325,
      "p95_ms": 7.884938377141952,
      "skipped": false,
      "error": null
    },
    {
      "repetition": 1,
      "source_url": "https://llm-speed.com/r/r_9suyze4n7lx",
      "ctx": 128000,
      "prompt_tokens_actual": 95212,
      "output_tokens": 96,
      "prefill_tps": 4428.074073061564,
      "decode_tps": 117.58587105022652,
      "ttft_ms": 21501.898664981127,
      "wall_ms": 22315.81017997861,
      "p50_ms": 8.401511505246162,
      "p95_ms": 8.57625944018364,
      "skipped": false,
      "error": null
    },
    {
      "repetition": 2,
      "source_url": "https://llm-speed.com/r/r_fzn_lluvw14",
      "ctx": 32000,
      "prompt_tokens_actual": 23872,
      "output_tokens": 98,
      "prefill_tps": 5891.147093639415,
      "decode_tps": 130.15110253900494,
      "ttft_ms": 4052.181963980198,
      "wall_ms": 4799.46634298563,
      "p50_ms": 7.511073976755142,
      "p95_ms": 7.556634999811649,
      "skipped": false,
      "error": null
    },
    {
      "repetition": 2,
      "source_url": "https://llm-speed.com/r/r_fzn_lluvw14",
      "ctx": 64000,
      "prompt_tokens_actual": 47652,
      "output_tokens": 100,
      "prefill_tps": 5478.670295233587,
      "decode_tps": 127.02932424243787,
      "ttft_ms": 8697.730915009975,
      "wall_ms": 9480.53266197443,
      "p50_ms": 7.807197496294975,
      "p95_ms": 7.866726246476174,
      "skipped": false,
      "error": null
    },
    {
      "repetition": 2,
      "source_url": "https://llm-speed.com/r/r_fzn_lluvw14",
      "ctx": 128000,
      "prompt_tokens_actual": 95212,
      "output_tokens": 96,
      "prefill_tps": 4414.945537614735,
      "decode_tps": 117.18462584586108,
      "ttft_ms": 21565.837944954634,
      "wall_ms": 22382.28950098157,
      "p50_ms": 8.413773506879807,
      "p95_ms": 8.54343496710062,
      "skipped": false,
      "error": null
    },
    {
      "repetition": 3,
      "source_url": "https://llm-speed.com/r/r_kdmexe0lkqn",
      "ctx": 32000,
      "prompt_tokens_actual": 23872,
      "output_tokens": 98,
      "prefill_tps": 5965.828013538053,
      "decode_tps": 129.81955925197605,
      "ttft_ms": 4001.4562849998474,
      "wall_ms": 4750.739704996347,
      "p50_ms": 7.50764499604702,
      "p95_ms": 7.818210773169994,
      "skipped": false,
      "error": null
    },
    {
      "repetition": 3,
      "source_url": "https://llm-speed.com/r/r_kdmexe0lkqn",
      "ctx": 64000,
      "prompt_tokens_actual": 47652,
      "output_tokens": 100,
      "prefill_tps": 5457.589352197361,
      "decode_tps": 127.12279138920348,
      "ttft_ms": 8731.327500998974,
      "wall_ms": 9513.464421987534,
      "p50_ms": 7.807701483368874,
      "p95_ms": 7.83894622027874,
      "skipped": false,
      "error": null
    },
    {
      "repetition": 3,
      "source_url": "https://llm-speed.com/r/r_kdmexe0lkqn",
      "ctx": 128000,
      "prompt_tokens_actual": 95212,
      "output_tokens": 96,
      "prefill_tps": 4420.309544219914,
      "decode_tps": 117.56070820542296,
      "ttft_ms": 21539.667991012335,
      "wall_ms": 22353.637015014887,
      "p50_ms": 8.406882494688034,
      "p95_ms": 8.470832452178001,
      "skipped": false,
      "error": null
    }
  ]
}
