{
 "version": "2026-11-18",
 "dense": {
  "pct": 9,
  "bias": -6,
  "n": 75,
  "models": 20
 },
 "moe-small": {
  "pct": 15,
  "bias": -7,
  "n": 88,
  "models": 15
 },
 "moe-large": {
  "pct": 34,
  "bias": -33,
  "n": 70,
  "models": 18
 },
 "method": "Median of |formula / measured - 1| over every measured plain-decode row (sourced or community) the speed formula can describe: a stated quant at 8 bits or fewer (not sub-2-bit), on a full-bandwidth chip, not flagged anchor:false, SSD-streamed or binned. The formula is applied alone, as for a model with no run of its own: dense = 1 / (params x GB-per-billion / (bandwidth x 0.7828) + 1.46 ms); small MoE (<= 40B total) the same on active parameters with 4.97 ms; large MoE 0.42 x the dense formula on active parameters. bias = the signed median (negative: the formula undershoots).",
 "maxBits": 8,
 "url": "https://llmcheck.net/methodology#typical-error"
}
