SemIf
16.1#183 of 261 on the open-weights board · Capability #165
Model guide · Formerly OpenJev · MIT logit reader on Qwen3.5-4B · updated October 26, 2026
Semantic ifs from a frozen open model, with no generation at all.
SemIf takes an unstructured state, criteria you write at request time and a set of typed options, runs one forward pass through an open model and reads the native logits for the option tokens. Nothing is generated, so there is no sentence to parse and no malformed JSON to retry. It was called OpenJev until it was renamed, which is why older benchmark tables and blog posts still use that name. The code is MIT; the weights keep their upstream licences.
SemIf is a third-party project by TheoLeeCJ. Jev AI does not host it; this page is a guide to what it is and how to run it yourself.
Try it on this page
Edit the example or paste your own text, then run it. Sign in to run: new accounts get 20 welcome credits, and a failed run costs nothing.
Jev AI does not host SemIf, so this editor runs Jev 1.14 on the same kind of decision. Compare the answers with your own SemIf setup.
Support ticket triage · Jev 1.14
1 Input
Hi, this is the third time I am writing. Our payouts have failed every day since Monday and my team cannot pay suppliers. I already re-entered the bank details twice. If this is not fixed today we will have to move to another provider.
2 Questions
One benchmark measured every system under one method. It ranks 261 open-weights systems on one board and 54 hosted APIs on another; Jev, a hosted API, is ranked on the second and shown on the first as an unranked reference. The full board is on the JevBench results page.
SemIf
16.1#183 of 261 on the open-weights board · Capability #165
Jev 1.13.0
71.4Open-weights reference row · #18 of 54 on the API board · Capability #15
IntelligenceHow often it picks the right answer, half from sealed decisions
CalibrationWhether 0.8 really means about 80%
SpeedMeasured response time
Open half 30.4 · sealed half 23.8; chance-corrected by request type: choice 47.6, yes/no 8.3, score 25.3. The composite also weighs a fourth axis, cost, which these bars leave out. Self-hosted latency is adjusted by the benchmark to approximate production load.
SemIf reproduces Jev’s interface pattern, not Jev’s undisclosed model or training: unstructured state, runtime criteria, typed options, and a probability for every option from one forward pass. The project is explicit about that distinction.
The speed argument is the strongest part of the project. On an RTX 3090 with a frozen state and 21 binary criteria, reading typed logits directly takes 1.023 seconds against 5.332 seconds to generate the same JSON array. With parallel suffix reuse the author measures 20.03 decisions per second against 2.33 for the autoregressive path.
On accuracy the author is careful: on a 102-row subset of TypeSafe’s cases, direct logits reach 0.845 modal agreement against Jev’s published 0.883, and balanced accuracy on authored decisions is 0.813. The README notes that this covers 102 aligned rows, not TypeSafe’s full 711-row aggregate, and asks you to calibrate and validate per workload.
JevBench v1.7.48 ranks SemIf on Qwen3.5-4B #183 of 261 open-weights systems at 16.1, measured on an RTX 5090. Speed is 90.4 and calibration 83.6; intelligence is 27.1, with 23.8 on the sealed half, and yes/no questions score 8.3 above chance. In v1.5.6 it was #14; the full re-measure on the new pool moved it down sharply. Jev, the board’s unranked reference, scores 71.4 with intelligence 63.6.
JevK5 builds on the same readout and base model, adds a distilled LoRA and a fitted temperature, and ranks #139; Plumb-4B, fine-tuned from JevK5, ranks #112. If you like SemIf’s approach but need more accuracy, that is the lineage to follow.
Clone the repository, install it against a GPU with the Qwen3.5-4B weights, and call it with a state, criteria and options. There is also a WebGPU demo that runs the smaller models in a browser tab.
git clone https://github.com/TheoLeeCJ/SemIf
cd SemIf
# Follow README.md to install and choose a model (Qwen3.5-4B is the primary baseline).
# A request supplies the state, runtime criteria and typed options; the response
# is a probability per option, read from the model's logits in one pass.From the project’s own documentation and JevBench v1.7.48.
The honest comparison is on your data, not on either scoreboard. Running Jev here takes a minute with no setup; if the answers on your awkward cases are close enough, SemIf is a legitimate way to run typed decisions on hardware you already own.
SemIf is the current name of the project formerly called OpenJev, by TheoLeeCJ. Other projects also use the openJev name, including razorback16’s DiffusionGemma server and openJev Verdict, and they are unrelated codebases by different authors.
An MIT-licensed project that reads typed option probabilities straight from the logits of a frozen open model such as Qwen3.5-4B, in one forward pass, with the state, criteria and options supplied at request time. It reproduces Jev’s interface, not Jev’s model.
In JevBench v1.7.48 it ranks #183 of 261 open-weights systems at 16.1 with intelligence 27.1 against Jev’s 63.6. On the author’s 102-row subset of TypeSafe’s cases it reaches 0.845 modal agreement against Jev’s published 0.883.
The published numbers are on an RTX 3090 with a 4B model. Qwen3-0.6B and MiniCPM5-2B run on less, and there is a WebGPU browser demo, but production throughput assumes a GPU you keep running.
It returns probabilities conditional on the options you supply and the README asks you to calibrate and validate per workload. JevBench v1.7.48 scores its calibration at 83.6 against Jev’s 90.6.
Every figure on this page is quoted from a published source and was read on October 26, 2026.
SemIf and every other product named on this page belongs to its respective owner and is not affiliated with Jev AI. Figures are quoted from the sources above as they read on October 26, 2026 and can change.