特点Highlights
一次前向，不生成文字One pass, no generation
直接读出每个选项的概率，不做逐字生成，速度快，结果稳定。Reads out option probabilities directly instead of generating text — fast and deterministic.
概率经过校准Calibrated probabilities
类别平衡的先验去偏加上按题型的温度缩放，输出的概率可以直接当作置信度使用。Class-balanced prior debiasing and per-type temperature scaling make the probabilities usable as confidence.
三种题型Three question types
支持 choice、noul 和 score 三种结构化题型，一个请求里可以同时问多个问题。Supports choice, noul and score questions — ask many at once in a single request.
问题之间互不影响Isolated questions
多个问题共享同一段上下文时只编码一次；增加或调换问题，不会改变其他问题的结果。A shared context is encoded once, and adding or reordering questions never changes another question's answer.
开箱即用Ready to run
一条命令用 Ollama 拉取运行；也可以用自带的 HTTP 服务，在 CUDA、Apple 芯片（MPS）或 CPU 上提供 /v1/systemone 接口。Pull and run it with Ollama in one command, or use the built-in server to expose /v1/systemone on CUDA, Apple silicon (MPS) or CPU.
完整可复现Fully reproducible
训练脚本、数据构建和 7 套基准测评全部开源，测评数据随仓库提供。Training, data building and all seven benchmark suites are open source, with the evaluation data included.
JevBench 对比JevBench comparison
| 模型Model | 参数量Params | 开放权重Open weights | 总分Overall |
|---|---|---|---|
| MiniCPM5-2B-Jev | 2.0B | ✓ | 78.8% |
| decider-2b v11 | 1.9B | ✓ | 76.2% |
| Kev-4B (r10) | 4.2B | ✓ | 75.8% |
| system-one-open | ~2B | ✓ | 73.2% |
| system-one (Qwen3-8B) | 8.2B | ✓ | 71.9% |
| Bespoke Nimble 9B | 9.0B | — | 67.5% |
* 在官方 JevBench（231 道决策题）上测评，对比模型取自 2026 年 9 月的公开排行榜。完整的 7 套基准结果见 GitHub。* Evaluated on the official JevBench (231 decisions); other models are from the public leaderboard as of September 2026. Full results across all seven suites are on GitHub.
快速开始Quickstart
用 Ollama 运行（最简单）With Ollama (easiest)
ollama pull actbro/minicpm5-2b-jev:2b
curl http://localhost:11434/v1/systemone -d '{
"model": "actbro/minicpm5-2b-jev:2b",
"state": "Ticket #402: Customer wants to cancel and get a full refund.",
"questions": {
"is_refund": { "type": "noul", "instructions": "Is the customer asking for a refund?" }
}
}'需要 Ollama 0.35 或更高版本，模型约 5 GB。更多用法（Python SDK、/api/generate 兼容模式）见 Ollama 页面。Requires Ollama 0.35 or later; the model is about 5 GB. See the Ollama page for the Python SDK and /api/generate compatibility mode.
从源码运行From source
git clone https://github.com/yuting-ai/minicpm5-2b-jev
cd minicpm5-2b-jev
pip install -r requirements.txt
python serve.py --checkpoint checkpoints/stage2 --port 8013启动后向 http://127.0.0.1:8013/v1/systemone 发送 JSON 请求即可。Python 调用、测评和训练方法见 README。Then send JSON requests to http://127.0.0.1:8013/v1/systemone. See the README for the Python API, evaluation and training.