SemIf runs open-source LLMs locally in the browser via WebGPU to compare two decision methods: reading option probabilities directly from logits versus autoregressively generating those probabilities as JSON tokens.

It loads quantized GGUF weights for Qwen3 0.6B, MiniCPM5 2B, and Qwen3.5 4B through wllama with no backend; model files and user inputs stay in the browser. Wall-time benchmarks for both methods are measured on the user's own GPU.