{
  "title": "GLM-5.3-Flash on a workstation with 16 GB of VRAM",
  "test_date": "2026-10-01",
  "prepared_date": "2026-10-02",
  "scope": "Summary of existing local evidence; no new benchmark was run for this article.",
  "model": "GLM-5.3-Flash",
  "pack": {
    "publisher": "Unsloth",
    "format": "UD-Q4_K_XL",
    "bytes": 199707321472,
    "revision": "621d456e93e926e4b52f85cff5f634358c1828f9",
    "shards": 6,
    "notes": "Mixed Q4/Q5/Q6 routed expert weights; most other matrices Q8. MTP and vision unused."
  },
  "hardware": {
    "cpu": "AMD Ryzen 9 9900X 12-Core Processor",
    "gpu": "NVIDIA GeForce RTX 5070 Ti",
    "gpu_reported_memory_mib": 16303,
    "installed_ram_bytes": 68719476736,
    "os_visible_ram_bytes": 66094223360,
    "os": "Microsoft Windows 11 Pro"
  },
  "placement": "llama.cpp CPU memory-mapped expert execution; GPU non-expert weights. No Strata adaptive Qwen expert scheduler.",
  "timing_scope": "fresh processes and same-process reset; filesystem caches uncontrolled; not cold-cache benchmarks",
  "code": {
    "prompt": "Write a Python function add(a, b) that returns a + b. Output code only.",
    "requests": [
      {
        "repetition": 0,
        "answer": "```python\ndef add(a, b):\n    return a + b\n```",
        "prompt_tokens": 31,
        "output_tokens": 17,
        "stop_token": 154827,
        "prefill_seconds": 44.00792880001245,
        "time_to_first_token_including_setup_seconds": 92.2417160999903,
        "decode_steps": 17,
        "decode_steps_per_second": 0.6480772691076129,
        "decode_latency_median_seconds": 1.5544668000075035,
        "decode_latency_min_seconds": 1.1339605000102893,
        "decode_latency_max_seconds": 1.9720626999915112
      },
      {
        "repetition": 1,
        "answer": "```python\ndef add(a, b):\n    return a + b\n```",
        "prompt_tokens": 31,
        "output_tokens": 17,
        "stop_token": 154827,
        "prefill_seconds": 47.53237159998389,
        "time_to_first_token_including_setup_seconds": null,
        "decode_steps": 17,
        "decode_steps_per_second": 0.6786721209871271,
        "decode_latency_median_seconds": 1.49366929999087,
        "decode_latency_min_seconds": 1.1533151000039652,
        "decode_latency_max_seconds": 1.8147759000130463
      }
    ],
    "memory": {
      "sampled_peak_bytes": {
        "rss": 52235214848,
        "private": 12749750272
      },
      "poll_seconds": 0.1,
      "os_peak_working_set_bytes": 52257955840,
      "os_peak_pagefile_bytes": 12752293888,
      "host_os_visible_bytes": 66094223360,
      "minimum_available_host_bytes": 263868416,
      "gpu_peak_used_MiB": 11064,
      "scope": "RSS includes mapped file pages; private commit is separate; GPU snapshots cover the whole device, including other apps, every ~5s"
    }
  },
  "greeting": {
    "prompt": "Reply with exactly: Hello Brian.",
    "requests": [
      {
        "repetition": 0,
        "answer": "Hello Brian.",
        "prompt_tokens": 19,
        "output_tokens": 4,
        "stop_token": 154827,
        "prefill_seconds": 25.873899500002153,
        "time_to_first_token_including_setup_seconds": 64.7041017000156,
        "decode_steps": 4,
        "decode_steps_per_second": 0.34423406139841434,
        "decode_latency_median_seconds": 1.5242139500041958,
        "decode_latency_min_seconds": 0.9480716999969445,
        "decode_latency_max_seconds": 7.62350100002368
      },
      {
        "repetition": 1,
        "answer": "Hello Brian.",
        "prompt_tokens": 19,
        "output_tokens": 4,
        "stop_token": 154827,
        "prefill_seconds": 13.728831699991133,
        "time_to_first_token_including_setup_seconds": null,
        "decode_steps": 4,
        "decode_steps_per_second": 1.4278556139574299,
        "decode_latency_median_seconds": 0.6660015499946894,
        "decode_latency_min_seconds": 0.4599451000103727,
        "decode_latency_max_seconds": 1.0094554999959655
      }
    ],
    "memory": {
      "sampled_peak_bytes": {
        "rss": 51610066944,
        "private": 12749660160
      },
      "poll_seconds": 0.1,
      "os_peak_working_set_bytes": 51642347520,
      "os_peak_pagefile_bytes": 12752203776,
      "host_os_visible_bytes": 66094223360,
      "minimum_available_host_bytes": 294952960,
      "gpu_peak_used_MiB": 11071,
      "scope": "RSS includes mapped file pages; private commit is separate; GPU snapshots cover the whole device, including other apps, every ~5s"
    }
  },
  "limitations": [
    "Two prescribed prompts, each reset and replayed.",
    "Decode rate counts calls including EOS and excludes prefill and setup.",
    "Whole-device GPU readings include other applications.",
    "Filesystem cache state was uncontrolled.",
    "No broad coding benchmark, long-context result, full higher-precision parity, or controlled disk-I/O profile."
  ],
  "outcome": "Generation passed the small functional checks. Project parked because coding speed was unsuitable for interactive use. The downloaded GLM pack was subsequently removed; code and evidence were retained."
}