Open Alternative to Jev

# Open Alternative to Jev

[](https://huggingface.co/spaces/IkerMoel/open-alternative-jev) [](LICENSE)

**Open-source System One models: typed, calibrated decisions from any open-weights LLM, in one forward pass.**
An open alternative to the idea behind TypeSafe's Jev, running on your own GPU with models you already have.
Python package `open-alternative-jev`, import name `so1` ("System One").
Also known as: open Jev, Jev alternative, open-source System One model.
This replaces the library, not the endpoint: it is a Python package you call in-process, and there is no HTTP server
or drop-in API for the official Jev SDK.

**Try it now:** [live demo on Hugging Face Spaces](https://huggingface.co/spaces/IkerMoel/open-alternative-jev) with Qwen3.5-4B, no install. No text is generated: the model reads the state once and every question is answered from the next-token distribution at its own position, restricted to the options you give.

## Results

Qwen3.6-27B (bitsandbytes 8-bit), one H200 MIG slice with 35 GB, Hugging Face Transformers. Every number
below comes from `benchmarks/results/`, produced by the scripts in `benchmarks/scripts/`.

**RACE-H, 250 passages x 4 questions (n = 1000).** One passage, four multiple-choice questions.
This is the shared-state case the library is built for.

| Mode | Accuracy | Questions / s | Tokens processed |
|---|---:|---:|---:|
| A: one question per forward | 92.6 % | 1.66 | 468,583 |
| B: batch of 4 (padding) | 92.8 % | 2.00 | 481,924 |
| **Open Alternative to Jev** (packed, state written once) | **92.9 %** | **4.55** | **186,898** |



Packing writes the passage once instead of four times: 2.5x fewer tokens, 2.3x the throughput of batching,
same accuracy (Open Alternative to Jev minus A = +0.3 points, 95 % CI -0.9 to +1.4, bootstrap over