SemIf: In-Browser LLM Decision Tool, No Backend
Original: OpenJev
Why This Matters
Demonstrates that structured LLM decision-making can run fully client-side, no API key or cloud backend required.
SemIf (formerly OpenJev) is an independent research project that runs local language models entirely in the browser to compare two decision-making methods: direct logit readout vs. JSON token generation. Three model sizes (0.6B to 4B) are supported, with no data sent to any server.
SemIf is a browser-only experiment that pits two inference approaches against each other using the same locally loaded model. The first method reads option probabilities directly from the model's logits, normalizing only across user-supplied choices without any decoding step. The second asks the model to write a full JSON distribution token by token — visible in real time. Both paths receive identical inputs, and wall-clock timing is measured with performance.now() so users see actual GPU performance, not canned results.
Three model tiers are available: Qwen3 0.6B (639 MB, aimed at phones), MiniCPM5 2B (1.56 GB, the desktop default), and Qwen3.5 4B (3.01 GB, high-memory desktops). Benchmark accuracy on the project's 102-row public subset ranges from 44% to 81.3% balanced accuracy, compared to the published Jev hosted value of 88.3%. Weights are pulled from Hugging Face, cached locally in the browser via pinned GGUF builds through wllama, and inputs never leave the page. The project is not affiliated with or endorsed by TypeSafe.