KOLDOS Local

Benchmark data format

The fields every benchmark result records.

A speed figure is only useful with the configuration behind it. Every result in the benchmark database records these fields.

Result shape

interface BenchmarkResult {
  id: string;
  gpuId: string | null;
  hardware: {
    gpu: string;
    vramGb: number | null;
    ramGb: number;
    cpu: string;
  };
  os: { id: "windows" | "linux" | "macos"; version: string };
  modelId: string;
  quantization: string;
  contextTokens: number;
  backend: { name: string; version: string };
  gpuLayers: "all" | number;
  tokensPerSecond: number | null;
  promptTokensPerSecond: number | null;
  generationTokensPerSecond: number | null;
  koldosVersion: string;
  measuredAt: string;
  source: "koldos-team" | "community";
  verified: boolean;
  notes: string | null;
}

Fields

Field	Type	Description
`gpuId`	string \| null	ID from the hardware database, or null for unlisted hardware.
`hardware.gpu`	string	Exact GPU model, including Laptop and memory variant.
`hardware.vramGb`	number \| null	Dedicated GPU memory in GB. Null for unified memory.
`hardware.ramGb`	number \| null	Installed system memory in GB. Null when it was not recorded.
`hardware.cpu`	string	CPU model as reported by the operating system.
`os`	{ id, version }	windows, linux or macos, plus the exact version when it was recorded.
`modelId`	string	ID of the KOLDOS model that was measured.
`quantization`	string	Quantization name exactly as published.
`contextTokens`	number	Context length configured for the run, in tokens.
`backend`	{ name, version }	Inference backend and its version.
`gpuLayers`	"all" \| number	How many layers were offloaded to the GPU.
`tokensPerSecond`	number \| null	End-to-end throughput: all tokens divided by total time.
`promptTokensPerSecond`	number \| null	Prompt processing speed.
`generationTokensPerSecond`	number \| null	Output generation speed.
`koldosVersion`	string \| null	Version of KOLDOS used for the run. Null for runs made before versions were recorded.
`measuredAt`	ISO 8601 date	When the run was recorded.
`source`	koldos-team \| community	Who ran the benchmark.
`verified`	boolean	Whether the KOLDOS team reproduced the result.

Rules for a comparable result

Name the exact GPU, including Laptop and memory variant.
Write the quantization exactly as published, for example Q4_K_M.
Use a fixed context length and record it.
Report prompt processing and generation speed separately. Leave a speed as null when it was not measured, never estimate it.
Only results reproduced by the KOLDOS team are marked as verified.