# SemIf

> Browser-based experiment comparing direct option probability readout with JSON generation on the same local language model, with timings measured on your GPU.

SemIf, formerly OpenJev, compares two ways to obtain probabilities for a decision's allowed options using a local language model. Users provide a state, question, and choices, then compare direct logit readout with JSON generation. It is an independent research project and states that it is not affiliated with or endorsed by TypeSafe.

## What It Is

The browser demo runs both methods sequentially on the same loaded model. Direct readout normalizes logits over the supplied option tokens without decoding; generation writes an estimated distribution as JSON. Inputs remain on the page, while model weights download from Hugging Face into the browser cache.

## Comparing Decision Methods

The interface reports setup, warmup, execution, first-token, and completion timings. Direct probabilities are conditional on the displayed options and are not calibrated confidence.

## Models and Browser Requirements

The available models are Qwen3 0.6B, MiniCPM5 2B, and Qwen3.5 4B. Published quality figures distinguish balanced accuracy from agreement on a public TypeSafe subset. Quantization can affect accuracy and speed. The demo requires WebGPU and an available GPU adapter; larger models require more memory and browser storage.

## Features
- Local browser inference with no backend
- Direct option probability readout without decoding
- JSON probability generation for comparison
- Sequential execution on the same loaded model
- Setup, warmup, first-token, and total timing measurements
- Choice of Qwen3 0.6B, MiniCPM5 2B, and Qwen3.5 4B
- Custom state, question, and allowed options
- Account support and email triage examples

## Integrations
Hugging Face, wllama

## Platforms
WEB

## Pricing
Paid

## Links
- Website: https://openjev.com/
- EveryDev.ai: https://www.everydev.ai/tools/semif