# JuL: jul-decision

> MiniCPM5-2B with a merged rank-16 LoRA and a Kev-style pointer head: a text and typed Choice, Noul or Score questions go in, a probability per option comes out of one forward pass, with no generated text. Built for JuL, a local library with a Jev-SDK-style client.

- Page: https://systemonemodels.ai/usejul/jul-decision
- API: https://api.systemonemodels.ai/v1/models/usejul/jul-decision
- Download: `pip install systemonemodels && systemone pull usejul/jul-decision`

## Facts

| | |
|---|---|
| Maker | JuL (https://systemonemodels.ai/usejul) |
| Decides | choice, score, noul |
| Architecture | jul-decision |
| Base model | openbmb/minicpm5-2b |
| Parameters | 2.5B |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-09-22 |
| Latest version | 2.0.0 |

## Reported evaluation

Suite: JuL hand-written bench (300 typed questions, 5 domains, 100 per type; the maker's own). Numbers are the publisher's own.

- Decision accuracy: 68.0%

- Median latency: 85 ms

## Model card

<!-- generated by scripts/seed_catalog.py; edit content/models/catalog.yaml -->

# jul-decision-minicpm5-2b

MiniCPM5-2B with a merged rank-16 LoRA and a Kev-style pointer head: a text and typed Choice, Noul or Score questions go in, a probability per option comes out of one forward pass, with no generated text. Built for JuL, a local library with a Jev-SDK-style client.

Brice Dauzats's open-source project. Trained with Kev's code (Jared Palmer, Apache-2.0) on about 30,000 human-written English and French typed decisions from commercially usable sources, per the card. Score is read by a vector reading that the jul library fits, not by the pointer head, and is the weakest type (0.41). The state is capped at 384 tokens and each branch at 1,024. No calibration error is reported; a temperature of 1.26 was fitted on held-out rows. The numbers are on the maker's own hand-written bench, not the third-party JevBench. Siblings: an MLX 4-bit build, jul-decision-wemm-4b (the library's default; 0.849 on the maker's 2,108-question bench) and jul-decision-e5-small (small enough for AWS Lambda). v2 weights landed on 5 October.

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **noul** — answers a yes/no question with one probability

## At a glance

| | |
|---|---|
| Parameters | 2.5B |
| Base model | `openbmb/MiniCPM5-2B` |
| Maker | JuL |
| Released | 2026-09-22 |
| License | apache-2.0 |
| Reported accuracy | 68.0% |
| Reported latency | 85 ms p50 per question, PyTorch bf16 on an NVIDIA A10G |

## Get the weights

```bash
pip install systemonemodels
systemone pull usejul/jul-decision
```

The files are served from the maker's Hugging Face repository, [`usejul/jul-decision-minicpm5-2b`](https://huggingface.co/usejul/jul-decision-minicpm5-2b), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/usejul/jul-decision-minicpm5-2b)
- [jul-decision-wemm-4b](https://huggingface.co/usejul/jul-decision-wemm-4b)
- [jul-decision-e5-small](https://huggingface.co/usejul/jul-decision-e5-small)
- [jul library](https://github.com/usejul/jul)

---

*This page was opened by System One for JuL, who can claim the organisation and take it over at any time.*

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models