Laya benchmark

Checks how fast Laya decides in this browser, and whether personality changes what it decides.

← Simul

1 · Load model

Checkpoint

Build · backend

WASM threads

Not loaded. The first load downloads the model from Hugging Face; after that it comes from the browser cache.

2 · Speed

Runs per batch size

Batch sizes (decisions per call)

Population

Each decision is one 5-option choice question over a ~120-token person + situation text, the realistic size for a person's state. Timing covers tokenizing plus inference, measured with performance.now(). The model runs on the main thread here, so the page stutters during calls; the simulator will run it in a Web Worker.

3 · Personality test

Same situation, four different people. If Laya understands personality, their answers should differ in ways that make sense.

Reverse option order (checks whether the model just favours the first option)

Results

Paste the JSON back to Claude to size the simulation.