idle intelligence
The Forecasting Company (TFC) t0-alpha ONNX INT8 export, run by onnxruntime-web (WebGPU with wasm fallback), next to the idle intelligence (II) t0-fast engine on GGUF weights (Q8_0/Q4_0), both loaded live in this tab. Be advised: a run downloads both models it compares, about 270 MB. Turning on the F32 reference below adds the never-quantized weights, about 680 MB total.
| engine | ep/backend | download MB | setup ms | isolated ms | warm median ms | warm p90 ms |
|---|
Setup is the first forecast after the model loads, so it includes compiling the GPU pipelines. Isolated is one forecast after the GPU has been idle for two seconds. Warm is the median of 10 forecasts, alternating between the two engines.
| engine | ms total | ms/signal | note |
|---|
| quantile | comparison | max-abs diff | max-abs % range | mean-abs diff | mean-abs % range |
|---|