User repository on GitHub
MiniCPM5-2B-Jev: High-Speed System 1 Decision Model (/v1/systemone, Jev-Compatible), #1 <=2B Model on JevBench
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Models & Reimplementations
Minicpm5 2B Jev
Open-weight 2B single-pass System One decision model on MiniCPM5-2B that answers a shared-state dictionary of choice, noul, and score questions in one forward pass without generating text.
- Stars
- 1
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- yuting-ai
- Last verified
- Oct 2026
About this project
MiniCPM5-2B-Jev is an open-source, single-pass, calibrated System 1 Decision Model built on OpenBMB's MiniCPM5-2B (weights on Hugging Face). Given a shared context document (state) and a dictionary of typed decision questions (choice, noul, score), it returns a calibrated probability distribution over the options for every question in one forward pass — without autoregressive text generation.
Quoted from the project's README. Read the full README.
Highlights
- Serves the structured systemone contract locally with benchmark tables and full per-suite results checked into the repo.
- Reads option letters straight from the base model's output head with debiasing and per-type temperature calibration.
- Reuses one shared-state encoding across questions with exact isolation, so siblings never shift each other's probabilities.
- Ships a small LoRA adapter plus calibration head, FastAPI server, unified evaluator, trainer, and dataset builder.
Watch out
Open weights with self-reported benchmarks; verify with the shipped evaluator. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
They are the repo's self-reported tables across JevBench, decision-v7, and other suites with results files checked in; rerun the unified evaluator yourself before treating any ranking as independent fact.
There is no autoregressive text to parse or schema to repair: one forward pass returns calibrated distributions for every question, which is cheaper and more stable for routing and gating code.
A Python install with the listed requirements, the shipped checkpoint loaded by device with bfloat16 where available, and enough accelerator memory for a 2B model plus LoRA adapter.
Score accuracy trails choice and noul on the headline decision-v7 and JevBench tables, so keep rubrics simple, calibrate thresholds on your own data, and escalate low-confidence scores.
Apache-2.0, built on OpenBMB's MiniCPM5-2B. The tuned weights are published on Hugging Face as ytbai/MiniCPM5-2B-Jev.
Shared-prefix KV-cache branching. When questions share a state of 256 tokens or more, the state is encoded once and each branch reuses the cached key-value tensors with pkv.crop, so adding or reordering sibling questions never changes another question's distribution.
About 25.3M, from LoRA with r=16 and alpha=32 attached across all 42 Transformer layers and all seven linear projections.
More Jev resources like this
6GitHub stars
Python helper that turns judgment calls into one-liners like likely, classify and rate, defaulting to Jev while also running on local or OpenAI-compatible backends.
Jev Models & Reimplementations#community#github
Community
Kungie/gutPython
Kungie/gutPython9GitHub stars
Self-hosted CPU-only server that speaks the /v1/systemone wire format for the KEV decision model, packaged as one container with ONNX Runtime and a health check.
Jev Models & Reimplementations#community#github
Community
aleskxyz/kev-onnxPython
aleskxyz/kev-onnxPython6GitHub stars
Laya multilingual server exposing POST /v1/systemone and GET /v1/models for noul, choice and score questions, usable with the official TypeSafe SDK.
Jev Models & Reimplementations#community#github
5GitHub stars
Python CLI that audits files and package trees against plain-English JSON rule packs in one System One forward pass, locally with kev or remotely with Jev.
Jev Models & Reimplementations#community#github
Community
skydiving94/klintPython
skydiving94/klintPython2GitHub stars
Learning project rebuilding a Jev-style decision model on a frozen Qwen3-0.6B backbone with a rank-16 LoRA, served alongside text generation from one weight set on Apple Silicon. Explicitly not production software.
Jev Models & Reimplementations#community#github
Community
DejaAI2/JevNextPython
DejaAI2/JevNextPython2GitHub stars
Self-hosted inference server for open Jev-like models behind POST /v1/systemone, starting with Clef-Flash on vLLM with micro-batching and Jev-compatible choice, score, and noul.
Jev Models & Reimplementations#community#github
Community
Arcobalneo/open-jevlike-inferPython
Arcobalneo/open-jevlike-inferPythonFrom the Jev guides
Original write-ups on what this entry is for.
Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.
Jev Models & Reimplementations#multimodal#open-models#decision-model
Read article
One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.
Jev Models & Reimplementations#decision-model#benchmarks#pricing
Read article
Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.
Jev Models & Reimplementations#open-models#hugging-face#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.