Skip to content

minicpm5-2b-jev

Open MiniCPM5-2B decision model serving /v1/systemone: typed choice, noul and score probabilities in one forward pass.

ModelresearchApache-2.0

Overview

MiniCPM5-2B-Jev is an open-weight, single-pass System 1 decision model built on OpenBMB's MiniCPM5-2B, with weights on Hugging Face. Given a shared state document and a dictionary of typed questions (choice, noul, score), it returns a probability distribution over each question's options without generating text, and it serves the same /v1/systemone request shape as Jev.

The README reports 78.8% overall on the public JevBench items, which it describes as first among open models of 2B parameters or fewer, against 86.6% for TypeSafe's hosted Jev 1.13.0 in the same table, plus 86.41% accuracy and 0.0307 ECE on a 10-source decision-v7 benchmark. It targets developers who want a small, self-hostable drop-in for Jev-style calls and researchers comparing open decision models.

What the decision model decides

TypeQuestionAnswers
ChoicePick one option for each typed question about the shared state—
Yes / noYes/no question about the stateyes · no
ScorePlace the state on an ordered rubric—

How it works

  1. Download weights from Hugging Face.
  2. Start the HTTP server exposing /v1/systemone.
  3. Send state plus a dictionary of typed questions.
  4. Read per-option probabilities and apply your own confidence threshold.

Strengths

  • Small 2B model that can be self-hosted.
  • Jev-compatible request contract, so client code can switch backends.
  • README reports low ECE (0.0307) on its decision-v7 benchmark.

Caveats

  • Benchmark figures are self-reported; the Hard split of JevBench remains well below hosted Jev.
  • decision-v7 is the author's own benchmark.

Getting started

Weights are on Hugging Face at ytbai/MiniCPM5-2B-Jev; the repo includes an HTTP server for the /v1/systemone contract.

At a glance

Type
Model
Status
research
Language
Python
Runtime
HTTP server
Has code
yes
License
Apache-2.0

Use cases and integrations

How curated lists describe it

Frequently asked questions

Is it compatible with Jev clients?

Yes, it natively serves the /v1/systemone structured decision contract.

How does it compare with hosted Jev?

Per its README table, 78.8% overall on public JevBench versus 86.6% for Jev 1.13.0.

Does it need a TypeSafe key?

No. The weights are open and run on your own hardware.

Related projects

  • Toy local Jev-shaped decision API on a 0.5B model that answers yes/no, choice and score questions by scoring option tokens.

    ServiceprototypePython
  • Small open decision models (0.8B and 4B, from Qwen3.5) with a server that speaks Jev's /v1/systemone format in one forward pass.

    ServicebetaPython
  • Serves MiniCPM5-2B-8bit on Apple Silicon as a Jev-compatible /v1/systemone service and reproduces its JevBench score.

    ModelprototypePython
  • reflex★ 166

    Small open decision model on a frozen Qwen3.5-4B serving /v1/systemone in about 200 ms, benchmarked against Jev on JevBench.

    ServiceresearchPython
  • Open Apache-2.0 Jev-like model family on Qwen3.8-27B that reads one-token candidate preferences and serves a TypeSafe-compatible API.

    ServiceresearchPython
  • Docker recipes and ARM64 images for serving ten native decision checkpoints, such as Cloudflare Clef, behind /v1/systemone on DGX Spark.

    ServicebetaPython