{
  "schema": "techfuel-dense-startup-timing/1",
  "case_id": "lmstudio-ollama-startup",
  "revision": "1.0.0",
  "status": "corrected",
  "observation_date": "2026-08-26",
  "credit": "TechFuelHQ, August 26, 2026; data: CC BY 4.0",
  "source_dataset": "https://techfuelhq.com/data/llm-server-compare/",
  "observations": "/data/lmstudio-ollama-startup-2026-08-26/observations.json",
  "scope": "Four normalized dense-model rows, on one RTX 5080. Unloaded-to-completed-short-reply, resident decode and resident server TTFT describe different events. Client-visible TTFT was not measured.",
  "change": "The original CSV's first request begins before model loading in Ollama and after a separate CLI load in LM Studio. This first versioned timing export states those boundaries and derives comparable short-reply estimates without changing the measurements.",
  "first_reply": "Servers already running; model unloaded; warm filesystem cache. One nonstreaming request per row, capped at 8 output tokens. On Ollama load_s is INSIDE the request; LM Studio load_wall_s is a separately timed CLI load BEFORE the request.",
  "resident": "Five short-prompt, 256-output-token requests per row. Decode seconds = median(n_pred / decode_tps). TTFT, when present, is server-reported, not client streaming latency.",
  "timings": [
    {
      "record_id": "ollama:llama3.2:3b",
      "revision": "1.0.0",
      "observation_date": "2026-08-26",
      "model": "Llama 3.2 3B",
      "app": "ollama 0.32.15",
      "model_id": "llama3.2:3b",
      "gguf_sha256": "dde5aa3fc5ffc17176b5e8bdc82f587b24b2678c6c66101bf7da77af9f7ccdff",
      "engine": "vendored llama.cpp 9d77fa172",
      "gpu": "NVIDIA GeForce RTX 5080",
      "driver": "610.88",
      "condition": "normalized",
      "context_tokens": 4096,
      "capture_at_utc": "2026-08-26T07:21:02.920621+00:00",
      "capture_sha256_lf": "2307f0d9fcfa2660877623a0a130edab0e79222b4fd12c1ee1224f21d8b4fd64",
      "load_s": 2.27,
      "request_s": 2.38,
      "request_includes_load": true,
      "unloaded_to_short_reply_approx_s": 2.38,
      "resident_decode_256_median_approx_s": 0.73,
      "resident_server_ttft_median_ms": null,
      "unloaded_requests": 1,
      "max_short_tokens": 8,
      "resident_runs": 5
    },
    {
      "record_id": "lms:llama-3.2-3b-q4km",
      "revision": "1.0.0",
      "observation_date": "2026-08-26",
      "model": "Llama 3.2 3B",
      "app": "LM Studio 0.4.21",
      "model_id": "llama-3.2-3b-q4km",
      "gguf_sha256": "dde5aa3fc5ffc17176b5e8bdc82f587b24b2678c6c66101bf7da77af9f7ccdff",
      "engine": "llama.cpp-win-x86_64-nvidia-cuda12-avx2@2.29.1",
      "gpu": "NVIDIA GeForce RTX 5080",
      "driver": "610.88",
      "condition": "normalized",
      "context_tokens": 4096,
      "capture_at_utc": "2026-08-26T07:15:29.298540+00:00",
      "capture_sha256_lf": "471f94d92c3025dd37cd83b056b706689a6e2e0c05c4d5564520c569ee5f4655",
      "load_s": 2.76,
      "request_s": 0.07,
      "request_includes_load": false,
      "unloaded_to_short_reply_approx_s": 2.83,
      "resident_decode_256_median_approx_s": 0.81,
      "resident_server_ttft_median_ms": 9.2,
      "unloaded_requests": 1,
      "max_short_tokens": 8,
      "resident_runs": 5
    },
    {
      "record_id": "ollama:qwen2.5:14b",
      "revision": "1.0.0",
      "observation_date": "2026-08-26",
      "model": "Qwen 2.5 14B",
      "app": "ollama 0.32.15",
      "model_id": "qwen2.5:14b",
      "gguf_sha256": "2049f5674b1e92b4464e5729975c9689fcfbf0b0e4443ccf10b5339f370f9a54",
      "engine": "vendored llama.cpp 9d77fa172",
      "gpu": "NVIDIA GeForce RTX 5080",
      "driver": "610.88",
      "condition": "normalized",
      "context_tokens": 4096,
      "capture_at_utc": "2026-08-26T07:21:02.920621+00:00",
      "capture_sha256_lf": "2307f0d9fcfa2660877623a0a130edab0e79222b4fd12c1ee1224f21d8b4fd64",
      "load_s": 5.52,
      "request_s": 5.68,
      "request_includes_load": true,
      "unloaded_to_short_reply_approx_s": 5.68,
      "resident_decode_256_median_approx_s": 2.5,
      "resident_server_ttft_median_ms": null,
      "unloaded_requests": 1,
      "max_short_tokens": 8,
      "resident_runs": 5
    },
    {
      "record_id": "lms:qwen2.5-14b-q4km",
      "revision": "1.0.0",
      "observation_date": "2026-08-26",
      "model": "Qwen 2.5 14B",
      "app": "LM Studio 0.4.21",
      "model_id": "qwen2.5-14b-q4km",
      "gguf_sha256": "2049f5674b1e92b4464e5729975c9689fcfbf0b0e4443ccf10b5339f370f9a54",
      "engine": "llama.cpp-win-x86_64-nvidia-cuda12-avx2@2.29.1",
      "gpu": "NVIDIA GeForce RTX 5080",
      "driver": "610.88",
      "condition": "normalized",
      "context_tokens": 4096,
      "capture_at_utc": "2026-08-26T07:17:23.614002+00:00",
      "capture_sha256_lf": "03c01f286ea7e1a849bcffc8f7dc2ebc2dc9f3d943cdd047f0cc808e61f1df6a",
      "load_s": 7.68,
      "request_s": 0.16,
      "request_includes_load": false,
      "unloaded_to_short_reply_approx_s": 7.84,
      "resident_decode_256_median_approx_s": 2.62,
      "resident_server_ttft_median_ms": 24.7,
      "unloaded_requests": 1,
      "max_short_tokens": 8,
      "resident_runs": 5
    }
  ]
}
