{
 "date": "2026-11-07",
 "model": "Kimi K3.1 Pro: 2.8T MoE, 92 MoE layers, 896 experts per layer, 16 routed per token; Q8_0 spine, MXFP4 experts (82,432 files, 1.45 TB)",
 "engine": "argonautlabsai/argodrive-ds4-kimi-k3.1, commit c0dcf692b4fb5f1abb05a942ca222c767b153351",
 "drafter": "Qwen/Qwen3-0.6B-Base, revision da87bfb608c14b7cf20ba1ce41287e8de496c0cd, float16 on the same GPU; 7 proposals plus the anchor make an 8-row block; gate 0.6",
 "hardware": "one RTX PRO 6000 Blackwell (96 GB) in a 70 GiB VM; RAID0 of three NVMe with 79,365 experts; second NVMe with 41,078; Thunderbolt NVMe with 74,000",
 "settings": "documented defaults plus a third expert directory, DS4_K3_TRACE_HASH=1, -c 2048, raw mode, greedy, one fresh process per run, monitoring off",
 "prompts": {
  "record": "The three main financial statements are",
  "public_512": "prompt-512-tokens.txt in the deltafin Kimi K3 benchmark"
 },
 "output_check": "token ids and FNV-1a hashes of every position's logits compared with the plain run of the same prompt; 5325 logit vectors compared across the final set and the qualification runs, 0 differ",
 "chat": "286-token chat prompt, text not published, 200 generated: plain 1.95 tok/s, gate 0.7 1.80, gate 0.9 1.90 with no block fired; the drafter is off for chat by default",
 "failure_tests": "forced rejection of the first, a middle and the last draft, a mix, end-of-text inside a block, a sidecar that dies, a malformed answer and a sidecar that cannot start: token ids and logits identical to plain in each",
 "not_established": [
  "comparison with Moonshot's own implementation",
  "exact match on other GPUs or drivers",
  "cancellation in the middle of a block and persistent-server recovery with the drafter on"
 ],
 "runs": [
  {
   "tag": "128P",
   "group": "record 128",
   "steady": 2.0,
   "steady_s": 63.41,
   "n": 127,
   "generated": 128,
   "wall": 93.18,
   "blocks": 0,
   "accepted": 0,
   "drafts": 0,
   "full": 0,
   "partial": 0,
   "zero": 0,
   "calls": 0,
   "call_ms": 0,
   "blk_tokens": 0,
   "ids": "reference: the other runs of this prompt are compared with this one",
   "hashes": "reference: the other runs of this prompt are compared with this one"
  },
  {
   "tag": "128Q",
   "group": "record 128",
   "steady": 2.34,
   "steady_s": 54.32,
   "n": 127,
   "generated": 128,
   "wall": 85.38,
   "blocks": 8,
   "accepted": 54,
   "drafts": 56,
   "full": 7,
   "partial": 1,
   "zero": 0,
   "calls": 66,
   "call_ms": 24.7,
   "blk_tokens": 62,
   "ids": "identical over 128",
   "hashes": "128 common, 0 differ"
  },
  {
   "tag": "128Qb",
   "group": "record 128",
   "steady": 2.34,
   "steady_s": 54.31,
   "n": 127,
   "generated": 128,
   "wall": 85.282,
   "blocks": 8,
   "accepted": 54,
   "drafts": 56,
   "full": 7,
   "partial": 1,
   "zero": 0,
   "calls": 66,
   "call_ms": 24.6,
   "blk_tokens": 62,
   "ids": "identical over 128",
   "hashes": "128 common, 0 differ"
  },
  {
   "tag": "128Pb",
   "group": "record 128",
   "steady": 2.02,
   "steady_s": 62.81,
   "n": 127,
   "generated": 128,
   "wall": 91.381,
   "blocks": 0,
   "accepted": 0,
   "drafts": 0,
   "full": 0,
   "partial": 0,
   "zero": 0,
   "calls": 0,
   "call_ms": 0,
   "blk_tokens": 0,
   "ids": "identical over 128",
   "hashes": "128 common, 0 differ"
  },
  {
   "tag": "200P",
   "group": "record 200",
   "steady": 1.98,
   "steady_s": 100.5,
   "n": 199,
   "generated": 200,
   "wall": 128.802,
   "blocks": 0,
   "accepted": 0,
   "drafts": 0,
   "full": 0,
   "partial": 0,
   "zero": 0,
   "calls": 0,
   "call_ms": 0,
   "blk_tokens": 0,
   "ids": "reference: the other runs of this prompt are compared with this one",
   "hashes": "reference: the other runs of this prompt are compared with this one"
  },
  {
   "tag": "200Q",
   "group": "record 200",
   "steady": 2.52,
   "steady_s": 78.86,
   "n": 199,
   "generated": 200,
   "wall": 109.802,
   "blocks": 17,
   "accepted": 117,
   "drafts": 119,
   "full": 16,
   "partial": 1,
   "zero": 0,
   "calls": 82,
   "call_ms": 29.7,
   "blk_tokens": 134,
   "ids": "identical over 200",
   "hashes": "200 common, 0 differ"
  },
  {
   "tag": "200Qb",
   "group": "record 200",
   "steady": 2.52,
   "steady_s": 78.81,
   "n": 199,
   "generated": 200,
   "wall": 109.613,
   "blocks": 17,
   "accepted": 117,
   "drafts": 119,
   "full": 16,
   "partial": 1,
   "zero": 0,
   "calls": 82,
   "call_ms": 29.7,
   "blk_tokens": 134,
   "ids": "identical over 200",
   "hashes": "200 common, 0 differ"
  },
  {
   "tag": "200Pb",
   "group": "record 200",
   "steady": 1.98,
   "steady_s": 100.41,
   "n": 199,
   "generated": 200,
   "wall": 129.002,
   "blocks": 0,
   "accepted": 0,
   "drafts": 0,
   "full": 0,
   "partial": 0,
   "zero": 0,
   "calls": 0,
   "call_ms": 0,
   "blk_tokens": 0,
   "ids": "identical over 200",
   "hashes": "200 common, 0 differ"
  },
  {
   "tag": "512P",
   "group": "record 512",
   "steady": 1.92,
   "steady_s": 265.52,
   "n": 511,
   "generated": 512,
   "wall": 293.927,
   "blocks": 0,
   "accepted": 0,
   "drafts": 0,
   "full": 0,
   "partial": 0,
   "zero": 0,
   "calls": 0,
   "call_ms": 0,
   "blk_tokens": 0,
   "ids": "reference: the other runs of this prompt are compared with this one",
   "hashes": "reference: the other runs of this prompt are compared with this one"
  },
  {
   "tag": "512Q",
   "group": "record 512",
   "steady": 2.61,
   "steady_s": 195.65,
   "n": 511,
   "generated": 512,
   "wall": 226.624,
   "blocks": 53,
   "accepted": 363,
   "drafts": 371,
   "full": 51,
   "partial": 2,
   "zero": 0,
   "calls": 144,
   "call_ms": 36.8,
   "blk_tokens": 416,
   "ids": "identical over 512",
   "hashes": "512 common, 0 differ"
  },
  {
   "tag": "512Qb",
   "group": "record 512",
   "steady": 2.61,
   "steady_s": 196.1,
   "n": 511,
   "generated": 512,
   "wall": 226.943,
   "blocks": 53,
   "accepted": 363,
   "drafts": 371,
   "full": 51,
   "partial": 2,
   "zero": 0,
   "calls": 144,
   "call_ms": 37.7,
   "blk_tokens": 416,
   "ids": "identical over 512",
   "hashes": "512 common, 0 differ"
  },
  {
   "tag": "512Pb",
   "group": "record 512",
   "steady": 1.92,
   "steady_s": 265.86,
   "n": 511,
   "generated": 512,
   "wall": 294.421,
   "blocks": 0,
   "accepted": 0,
   "drafts": 0,
   "full": 0,
   "partial": 0,
   "zero": 0,
   "calls": 0,
   "call_ms": 0,
   "blk_tokens": 0,
   "ids": "identical over 512",
   "hashes": "512 common, 0 differ"
  },
  {
   "tag": "OP",
   "group": "public 512-token prompt, 200 generated",
   "steady": 1.58,
   "steady_s": 125.97,
   "n": 199,
   "generated": 200,
   "wall": 179.842,
   "blocks": 0,
   "accepted": 0,
   "drafts": 0,
   "full": 0,
   "partial": 0,
   "zero": 0,
   "calls": 0,
   "call_ms": 0,
   "blk_tokens": 0,
   "ids": "reference: the other runs of this prompt are compared with this one",
   "hashes": "reference: the other runs of this prompt are compared with this one"
  },
  {
   "tag": "OQ",
   "group": "public 512-token prompt, 200 generated",
   "steady": 2.29,
   "steady_s": 87.06,
   "n": 199,
   "generated": 200,
   "wall": 143.834,
   "blocks": 24,
   "accepted": 168,
   "drafts": 168,
   "full": 24,
   "partial": 0,
   "zero": 0,
   "calls": 25,
   "call_ms": 65.4,
   "blk_tokens": 192,
   "ids": "identical over 200",
   "hashes": "200 common, 0 differ"
  },
  {
   "tag": "OQb",
   "group": "public 512-token prompt, 200 generated",
   "steady": 2.29,
   "steady_s": 86.85,
   "n": 199,
   "generated": 200,
   "wall": 142.621,
   "blocks": 24,
   "accepted": 168,
   "drafts": 168,
   "full": 24,
   "partial": 0,
   "zero": 0,
   "calls": 25,
   "call_ms": 67.1,
   "blk_tokens": 192,
   "ids": "identical over 200",
   "hashes": "200 common, 0 differ"
  },
  {
   "tag": "OPb",
   "group": "public 512-token prompt, 200 generated",
   "steady": 1.58,
   "steady_s": 125.8,
   "n": 199,
   "generated": 200,
   "wall": 179.444,
   "blocks": 0,
   "accepted": 0,
   "drafts": 0,
   "full": 0,
   "partial": 0,
   "zero": 0,
   "calls": 0,
   "call_ms": 0,
   "blk_tokens": 0,
   "ids": "identical over 200",
   "hashes": "200 common, 0 differ"
  }
 ]
}