open-alternative-jev
Python library that gets typed, calibrated decisions from any open-weights LLM in one forward pass, with packed shared-state questions.
Overview
Open Alternative to Jev (package open-alternative-jev, import so1) provides System One-style typed decisions from any open-weights LLM via Hugging Face Transformers or vLLM, running on your own GPU with no training. The model reads the state once, and each question is answered from the next-token distribution at its own position, restricted to the options you supply. It is an in-process library, not an HTTP server or drop-in for the Jev SDK.
The README reports that on LocalLLaMA/typed-decisions (400 cases, 2,000 decisions), stock Qwen3.6-27B through the library scored 73.7% accuracy, ECE 0.020 and 582 ms per case, versus 72.7%, ECE 0.144 and 710 ms for Jev 1.13.0 as measured by the benchmark authors. Packing several questions over one written-once state on RACE-H gave the same accuracy with 2.5x fewer tokens and 2.3x the throughput of batching. A Hugging Face Space demo runs Qwen3.5-4B. It suits teams with GPUs who want local decisions.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which option answers each question about the shared state? | — |
| Yes / no | Yes/no question about the state | yes · no |
| Score | Score on supplied levels | — |
How it works
- Load an open model with Transformers or vLLM.
- Pack the state once followed by each question.
- Read each question's next-token distribution over its options.
- Use the calibrated probabilities in application logic.
Strengths
- No training required; works with models you have.
- Packed state cuts tokens and raises throughput.
- Reported calibration better than Jev on the shared benchmark.
Caveats
- Needs substantial GPU memory for large models.
- Library only; no Jev-compatible HTTP endpoint.
Getting started
pip install open-alternative-jev, then import so1 and point it at a Hugging Face or vLLM model.At a glance
- Type
- Library
- Status
- beta
- Language
- Python
- Runtime
- library (Transformers, vLLM)
- Has code
- yes
- Has tests
- yes
- Stars
- 67
Use cases and integrations
How curated lists describe it
Open alternative: runs a Jev-shaped decision model locally on your own GPU.
yibie/awesome-jevOpen-source alternative to TypeSafe's Jev: a System One style model layer that gives typed, calibrated decisions from any open-weights LLM in one forward pass (HF + vLLM), with honest benchmarks
wh000wh000/awesome-jev-liveOpen alternative to Jev: typed, calibrated decisions from any open-weights LLM in one forward pass (HF + vLLM), with benchmarks.
hellogumbo/awesome-jevTyped, calibrated decisions from any open-weights model in one forward pass, as the Python package `open-alternative-jev`.
MrJev/awesome-jev
Frequently asked questions
Does it beat Jev?
On LocalLLaMA/typed-decisions the README reports 73.7% vs 72.7% with much lower ECE, using Qwen3.6-27B.
Is it a drop-in Jev server?
No, it is an in-process Python library with no HTTP server.
Do I need to fine-tune?
No, it uses stock open-weight models.
Related projects
llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++MLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.
ProjectprototypePythonLocal server (xks) implementing Jev's /v1/systemone contract on an open-weight LLM, accelerated with Xeon Phi cards.
ProjectresearchRustEducational C11 library and CLI that turns DiffusionGemma 26B into a CPU-first OpenJev-style decision engine reading candidate logits.
ProjectresearchCExperimental engine that makes an existing LLM do Jev-like classification and scoring by reusing context and building JSON.
ProjectresearchPythonReads typed decisions from one forward pass of PrismML's 1.75-bit Bonsai 27B under llama.cpp, using the OpenJev/SemIf logit method.
ProjectresearchPython