Open Bonsai Jev

An independent implementation of the OpenJev mechanism running on PrismML's low-bit Bonsai weights: a 27B model in a 5.95 GB file that returns typed decisions from a single forward pass, with a score.sh CLI that reads probabilities off the next-token distribution.

It belongs to the same category: System One style models that return typed probabilistic decisions instead of text. The README frames hosted Jev as closed, waitlisted and vendor-owned, and this repo as a runnable open study of the same idea built on SemIf's mechanism.

The model is handed a lettered multiple-choice question and the scorer reads the probability of A versus B from the next-token distribution. Nothing is generated and nothing is parsed: one forward pass produces a real number an if-statement can branch on.

The README says it runs on CPU, Metal, CUDA or Vulkan from a 5.95 GB weights file on a laptop. The sample run reports a llama.cpp backend and a 1.852 second forward pass, and the benchmarks were produced on an Apple M4 Pro GPU with a 4096 context.

The badges headline WANLI-256 at 74.6% for the 5.95 GB 27B; the visible table shows the 1.7B at 52.0% and the 4B higher, against a label-balanced chance rate of 33.3%. The repo also mentions comparisons with hosted Jev on the same eval files, a 110-pair browser-agent benchmark, calibration and image evaluations.

The code is MIT and the Bonsai weights are Apache 2.0, used exactly as published with nothing retrained or requantized. The repo credits TheoLeeCJ (SemIf/OpenJev) and PrismML, and states it is not affiliated with, endorsed by or connected to TypeSafe, PrismML or TheoLeeCJ.

Not directly. openjev/score.py is an independent re-implementation of SemIf's direct mode with its own prompt and no vendored SemIf code, so the README warns the numbers cannot be compared one-to-one. For the reference implementation of the idea it points readers to SemIf.