{
  "date": "2026-09-11",
  "model": "GLM-5.3 full, 744B MoE; UD-Q4_K_XL base, routed experts requantized to uniform Q4_K",
  "model_bytes": 434170891168,
  "model_md5": "29ef5dbdad788cf4bf5d58f2ab129ea4",
  "upstream_base": "6289c51",
  "engine_binary_md5": "760516853f630561a65ff6989aacd324",
  "hardware": "M5 Max 128 GB; internal plus three enclosures",
  "prompt": "The three main financial statements are",
  "prompt_tokens": 24,
  "mode": "raw completion, greedy, no thinking",
  "metric": "Harness decode: generated tokens after the first response divided by elapsed time from first to last response; prefill/startup excluded",
  "comparison": "Same patched ds4 engine and four-drive layout; control disables poll-gated pre-commit and dense eviction scan; split reads and prefetch retained",
  "replica_weights": "10:5:5:4",
  "read_threads": 48,
  "dashboard": "Raw arm metadata records dashboard open; preserved here rather than describing these arms as monitor-free.",
  "summary": [
    {
      "generated_tokens": 128,
      "candidate": [
        4.213,
        4.2215,
        4.2134
      ],
      "control": [
        3.6602,
        3.6581,
        3.6612
      ],
      "candidate_median": 4.2134,
      "control_median": 3.6602,
      "gain_pct": 15.113928200644766
    },
    {
      "generated_tokens": 512,
      "candidate": [
        4.0161,
        3.9982,
        4.0578
      ],
      "control": [
        3.461,
        3.4688,
        3.4416
      ],
      "candidate_median": 4.0161,
      "control_median": 3.461,
      "gain_pct": 16.03871713377636
    }
  ],
  "runs": [
    {
      "arm": "qual128-cand-1",
      "variant": "candidate",
      "repetition": 1,
      "output_md5": "da46fa88983a2d172ad8c4c83f7382ae",
      "source_log_sha256": "162fade44821ed8f1b58244cde1dcf0bb737e0caf98748f3080216eb5e719512",
      "tokens_req": 128,
      "chunks": 128,
      "decode_tok_s": 4.213,
      "gen_s": 30.14,
      "first_byte_s": 4.51,
      "ds4_gen_tps": 4.25,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual128-ctrl-1",
      "variant": "control",
      "repetition": 1,
      "output_md5": "da46fa88983a2d172ad8c4c83f7382ae",
      "source_log_sha256": "b86f89fc1ed86758398ef6f40af7e00c2061370a099cc224cf14b109efa3a77a",
      "tokens_req": 128,
      "chunks": 128,
      "decode_tok_s": 3.6602,
      "gen_s": 34.7,
      "first_byte_s": 4.56,
      "ds4_gen_tps": 3.69,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual128-cand-2",
      "variant": "candidate",
      "repetition": 2,
      "output_md5": "da46fa88983a2d172ad8c4c83f7382ae",
      "source_log_sha256": "8040a0676924172a4b2c1ca8197c6eb7aec2a9bcf156ac941a388cbb2ba596a2",
      "tokens_req": 128,
      "chunks": 128,
      "decode_tok_s": 4.2215,
      "gen_s": 30.08,
      "first_byte_s": 4.44,
      "ds4_gen_tps": 4.25,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual128-ctrl-2",
      "variant": "control",
      "repetition": 2,
      "output_md5": "da46fa88983a2d172ad8c4c83f7382ae",
      "source_log_sha256": "c80bba809d30576e7c86f1f2256c2ba9a504e3a3ad62947e1e6a0c598f64eb9c",
      "tokens_req": 128,
      "chunks": 128,
      "decode_tok_s": 3.6581,
      "gen_s": 34.72,
      "first_byte_s": 4.51,
      "ds4_gen_tps": 3.69,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual128-cand-3",
      "variant": "candidate",
      "repetition": 3,
      "output_md5": "da46fa88983a2d172ad8c4c83f7382ae",
      "source_log_sha256": "cc0cf4dbc5941cd815ba07850aa6946fe9a830d0b68eaca15f339ffade97d83a",
      "tokens_req": 128,
      "chunks": 128,
      "decode_tok_s": 4.2134,
      "gen_s": 30.14,
      "first_byte_s": 4.56,
      "ds4_gen_tps": 4.25,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual128-ctrl-3",
      "variant": "control",
      "repetition": 3,
      "output_md5": "da46fa88983a2d172ad8c4c83f7382ae",
      "source_log_sha256": "24ad3c9cbd0e956b006d20d279b3fed5a7f87b4262fddf85bb5affd4c72392ca",
      "tokens_req": 128,
      "chunks": 128,
      "decode_tok_s": 3.6612,
      "gen_s": 34.69,
      "first_byte_s": 4.7,
      "ds4_gen_tps": 3.69,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual512-cand-1",
      "variant": "candidate",
      "repetition": 1,
      "output_md5": "6cc7187b3673eace94c49cb474606226",
      "source_log_sha256": "627dfa994c2d1f4ad8025fa86ead53cf9ab33404e65f24f0a956cb2d072198c5",
      "tokens_req": 512,
      "chunks": 512,
      "decode_tok_s": 4.0161,
      "gen_s": 127.24,
      "first_byte_s": 4.61,
      "ds4_gen_tps": 4.02,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual512-ctrl-1",
      "variant": "control",
      "repetition": 1,
      "output_md5": "6cc7187b3673eace94c49cb474606226",
      "source_log_sha256": "235d81fb585098ffe4bdf721e340c2bcf0fb7bd8a94979aca4121b315e30a566",
      "tokens_req": 512,
      "chunks": 512,
      "decode_tok_s": 3.461,
      "gen_s": 147.65,
      "first_byte_s": 4.63,
      "ds4_gen_tps": 3.47,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual512-cand-2",
      "variant": "candidate",
      "repetition": 2,
      "output_md5": "6cc7187b3673eace94c49cb474606226",
      "source_log_sha256": "50af667c279c19203f623964276bdcb2542a835c92c76163dd4cc809e7117041",
      "tokens_req": 512,
      "chunks": 512,
      "decode_tok_s": 3.9982,
      "gen_s": 127.81,
      "first_byte_s": 4.59,
      "ds4_gen_tps": 4.01,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual512-ctrl-2",
      "variant": "control",
      "repetition": 2,
      "output_md5": "6cc7187b3673eace94c49cb474606226",
      "source_log_sha256": "c3cdc96a373fe41b5c5464985bce1851a4dbe92537454d6ea31aeb8852e0ba28",
      "tokens_req": 512,
      "chunks": 512,
      "decode_tok_s": 3.4688,
      "gen_s": 147.31,
      "first_byte_s": 4.61,
      "ds4_gen_tps": 3.48,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual512-cand-3",
      "variant": "candidate",
      "repetition": 3,
      "output_md5": "6cc7187b3673eace94c49cb474606226",
      "source_log_sha256": "de1d3d2af9bc0fb6b4efb75a1558f5a8ec04bb6fd5b82e13af69be141bf403a0",
      "tokens_req": 512,
      "chunks": 512,
      "decode_tok_s": 4.0578,
      "gen_s": 125.93,
      "first_byte_s": 4.53,
      "ds4_gen_tps": 4.07,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    },
    {
      "arm": "qual512-ctrl-3",
      "variant": "control",
      "repetition": 3,
      "output_md5": "6cc7187b3673eace94c49cb474606226",
      "source_log_sha256": "58b2f308924ee9efeb8a13f3690a735373371c7ecf971aa03af71182cb2a66e1",
      "tokens_req": 512,
      "chunks": 512,
      "decode_tok_s": 3.4416,
      "gen_s": 148.48,
      "first_byte_s": 4.63,
      "ds4_gen_tps": 3.45,
      "swap_growth_mb": 0.0,
      "valid": "ok",
      "dashboard": "open",
      "rc": 0
    }
  ],
  "notes": [
    "Three interleaved candidate/control repetitions per length; identical output MD5 across each group of six.",
    "Requantization quality versus original quantization and FP8 reference was not established by this throughput campaign.",
    "A 12 September candidate-only recheck measured 3.8609 tok/s median at 512 tokens; this dated comparison does not claim a fresh September 14 result.",
    "These figures are separate from the DeepSeek V4.1 benchmark: different model, prompt, timing metric and baseline."
  ]
}
