# SemIf SemIf ("semantic ifs from open models", formerly called OpenJev) is an independent, MIT-licensed reimplementation of the [[Jev]] *interface* on top of open-weight models. It was built by TheoLeeCJ and isn't affiliated with [[TypeSafe AI]]. The trick is simple: instead of asking a model to write its answer as JSON and parsing it, SemIf does a single forward pass and reads the logits of the option tokens directly. On a frozen [[Qwen]]3.5-4B, that's 21 probability pairs in about 1 second on an RTX 3090, versus 5.3 seconds for autoregressive JSON (about 5.2x faster). Reusing the shared state across several decisions pushes throughput up about 20x. Quality trails the real thing: 0.813 balanced accuracy on the author's decisions versus 0.883 for Jev (the WebGPU page reports 0.845 modal agreement with Jev). And the author is clear about the limits: they did NOT reproduce Jev's model or training, so there are no calibration guarantees. What I like about it is the browser lab at openjev.com. It runs entirely on your device (Qwen3 0.6B on phones, MiniCPM5 2B by default, Qwen3.5 4B on high-memory machines) and shows both methods side by side with real timings. It's the fastest way to *feel* why logit reading beats text generation for decisions. It hit the front page of [[Hacker News]] (719 points, ~300 comments). ## References - [SemIf on GitHub](https://github.com/TheoLeeCJ/SemIf) - [openjev.com (browser lab)](https://openjev.com/) - [OpenJev on Hacker News](https://news.ycombinator.com/item?id=49752041) ## Related - [[Jev]] - [[System One Models]] - [[Kev]] - [[Jevlike]] - [[AI Open Weight Models]]