SemIf

About

SemIf is a browser-based experiment that runs small open language models (Qwen3, MiniCPM5) locally through WebGPU to compare two ways of making a constrained decision: reading raw option logits directly versus having the model generate the same probability distribution as JSON text token by token. It measures accuracy and timing for both methods entirely on the user's own GPU, with no backend and no data leaving the browser. It was previously published under the name OpenJev.