jev-open-source
Open-weights Jev-style decision engine on Qwen2.5-1.5B with a decision head and QLoRA, answering Choice, Noul and Score in one pass.
Overview
JEV Open Source is an open-weights System One decision engine inspired by a 2026 paper evaluating Jev. It builds on Qwen2.5-1.5B-Instruct, removes the causal language-model head and adds a dual-mode decision head trained with 4-bit QLoRA, so the model answers structured Choice, Noul and Score questions in a single forward pass instead of generating text.
It targets developers who use LLMs for operational micro-decisions such as routing queries, verifying grounding against retrieved documents, screening prompts and grading responses, and who want to avoid token-by-token latency, JSON parsing failures and uncalibrated outputs. The README claims 15-35 ms per query and a VRAM footprint under 2 GB, and covers dataset preparation, training, evaluation metrics, a Python inference API, CLI workflows and tests. These performance figures are the project's own claims.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which sub-agent should handle this query?(inferred) | — |
| Yes / no | Is this answer grounded in the retrieved documents?(inferred) | yes · no |
| Score | How good is this response against the rubric?(inferred) | — |
How it works
- Prepare and verify decision datasets
- Fine-tune the decision head with 4-bit QLoRA on Qwen2.5-1.5B
- Evaluate accuracy and calibration
- Serve decisions through the Python API or CLI
Strengths
- Runs locally on modest GPUs (under 2 GB VRAM claimed)
- Single forward pass with no syntax failures
- Open weights and training pipeline
Caveats
- Latency and calibration claims are self-reported
- Small base model limits accuracy on hard decisions
- Licence not stated in the extracted README
Getting started
Follow the Installation and Training sections; the base model is Qwen/Qwen2.5-1.5B-Instruct.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch model and CLI
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Open-source, non-autoregressive System One decision engine for low-latency…
wh000wh000/awesome-jev-live
Frequently asked questions
Does this use TypeSafe's Jev?
No. It is an independent open-weights model built on Qwen2.5-1.5B.
What hardware does it need?
The README claims under 2 GB of VRAM with 4-bit quantisation.
How fast is it?
The README claims 15-35 ms per query.
Related projects
Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Three open 4B decision models for English, Turkish and German that serve the Jev /v1/systemone API on a CPU in about 3 GB of RAM.
ServicebetaPython