{
  "schema": "cortiq-mobile-ui-throughput/v1",
  "date": "2026-10-01",
  "device": {
    "model": "iPhone 16",
    "soc": "A18"
  },
  "app": {
    "channel": "TestFlight",
    "version": "1.3.0",
    "build": 47
  },
  "model": {
    "repo": "infosave/LFM2.5-cmf",
    "file": "lfm2.5-230m-q4tp.cmf",
    "file_bytes": 132082125,
    "hub_lfs_sha256": "889fa1e3a133004f9fab273e24a355b51056c5e349f9efd4552892a440c42f9a",
    "quantization": "Q4TP",
    "layers": 14,
    "context_tokens": 128000,
    "estimated_ram_mib": 622
  },
  "settings": {
    "temperature": 0.1,
    "top_p": 0.95,
    "max_tokens": 128,
    "cpu_threads": "auto (5)",
    "gpu_toggle": false
  },
  "method": {
    "surface": "phone chat UI",
    "prompt": "Explain how CMF task masks work",
    "prompt_tokens_ui": 7,
    "warmup_excluded": true,
    "measurement": "UI-reported completion tokens / full execution interval measured by the engine after generation starts; prompt processing and model response included, queue wait excluded; UI display is rounded",
    "timing_precision": "UI-rounded values",
    "exploratory_other_models_note": "Their completion counts exceed the controlled LFM max_tokens cap; exact generation cap and sampling settings were not recorded, so these are not directly comparable with LFM."
  },
  "runs": [
    {
      "id": 0,
      "state": "warmup",
      "prompt_tokens": 7,
      "completion_tokens": 128,
      "latency_s_display": 0.8,
      "tokens_per_second_ui": 168.4
    },
    {
      "id": 1,
      "state": "warm",
      "prompt_tokens": 7,
      "completion_tokens": 37,
      "latency_s_display": 0.3,
      "tokens_per_second_ui": 144.5
    },
    {
      "id": 2,
      "state": "warm",
      "prompt_tokens": 7,
      "completion_tokens": 37,
      "latency_s_display": 0.3,
      "tokens_per_second_ui": 145.7
    },
    {
      "id": 3,
      "state": "warm",
      "prompt_tokens": 7,
      "completion_tokens": 125,
      "latency_s_display": 0.8,
      "tokens_per_second_ui": 166.0
    }
  ],
  "warm_aggregate": {
    "range_tokens_per_second_ui": [
      144.5,
      166.0
    ],
    "median_tokens_per_second_ui": 145.7,
    "mean_tokens_per_second_ui": 152.07,
    "runs_over_60_tokens_per_second": 3,
    "runs_total": 3
  },
  "additional_ui_capture": {
    "prompt": "Write a SQL query for monthly revenue",
    "prompt_tokens": 9,
    "completion_tokens": 95,
    "latency_s_display": 0.6,
    "tokens_per_second_ui": 162.1,
    "asset": "assets/15_lfm230_iphone_speed.png",
    "note": "Interactive evidence capture; not included in the controlled aggregate because the prompt differs."
  },
  "exploratory_other_models": [
    {
      "model": "bonsai-1.7b-q1.cmf",
      "gpu_toggle": false,
      "completion_tokens": 541,
      "latency_s_display": 11.3,
      "tokens_per_second_ui": 47.9,
      "prompt": "Explain how CMF task masks work",
      "series": "exploratory; separate configuration from the controlled LFM runs",
      "inference_mode": "Chat UI with generic ChatML fallback; shipped CMF header has no tokenizer_config.chat_template",
      "generation_settings": "not recorded; separate from controlled LFM settings"
    },
    {
      "model": "bonsai-1.7b-q1.cmf",
      "gpu_toggle": true,
      "completion_tokens": 261,
      "latency_s_display": 13.6,
      "tokens_per_second_ui": 19.2,
      "prompt": "Explain how CMF task masks work",
      "series": "exploratory; separate configuration from the controlled LFM runs",
      "note": "GPU toggle was enabled before reload; execution backend was not traced.",
      "inference_mode": "Chat UI with generic ChatML fallback; shipped CMF header has no tokenizer_config.chat_template",
      "generation_settings": "not recorded; separate from controlled LFM settings"
    }
  ],
  "restored_phone_state": {
    "active_model": "MiniCPM5-2B-q4tp",
    "temperature": 0.7,
    "top_p": 0.95,
    "max_tokens": 1024,
    "gpu_toggle": false,
    "language": "system (Russian)"
  },
  "limits": [
    "Not a thermal, battery, long-context, or quality evaluation.",
    "Result is specific to the device, build, model, settings, and prompt.",
    "Do not generalize the LFM result to every model in the catalog."
  ]
}