V
vnmartinez/jev-local-alternatives
Notes on TypeSafe AI's Jev decision model and a local open-source zero-shot alternative (GLiClass, CPU)
在 GitHub 查看 ↗链接已于 2026-11-12 核查 · HTTP 200 · 打开的是厂商自己的官网
安装与配置
取自该仓库的安装命令
pip install torch
解析自我们于 2026-10-26 抓取的该仓库 README；我们未执行过。
这条是仓库自己写的安装方式，不是 MCP 客户端配置，所以不生成 JSON。
配置项的键名以你的客户端为准：客户端与 CLI
- ★ 数取值
- 0 ★ · vnmartinez/jev-local-alternatives · 取自 GitHub 搜索 API · （2026-11-13 20:00:50 核对）
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型：开源项目
- 收录与活跃
- 收录于 2026-10-26 · 最近见到 2026-11-14 · 共见到 20 次 · 被 1 个来源列出
- 安装
pip install torch
常见问题（答案都来自我们的记录）
怎么安装？
该仓库给出的安装命令是
pip install torch。我们的记录里它是仓库自己提供的；不保证在你的客户端上直接可用。它适合谁 / 用来做什么？
我们还没有为它记录标签。
谁维护它、还在更新吗？
仓库
vnmartinez/jev-local-alternatives；我们收录于 2026-10-26；最近见到 2026-11-14。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。# jev-local-alternatives
以下摘录自该仓库的 README（片段） · 在 GitHub 读完整 README →
Notes on Jev, TypeSafe AI's "System One" decision model, plus a runnable example of an open-source model you can run locally for the same kind of task.
> Research snapshot from October 2026. Jev performance and adoption figures are self-reported by TypeSafe. No independent benchmark or architecture paper existed at the time of writing.
## What Jev is
A discriminative model. It does not generate text. You give it unstructured context and typed questions (boolean, choice, ordered scale), and it returns a probability distribution with a confidence score for each.
Labels are supplied at inference time (zero-shot), so it needs no task-specific training.
Reportedly trained only on synthetic data with an unpublished method (RLCD) that targets calibrated probabilities. TypeSafe says it is transformer-based. Weights are closed.
Released 15 Sep 2026 as a hosted API only. There is no self-hosting.
Listed pricing: US$0.042 per 1M input tokens, output free. Claimed latency is 70–500 ms. Input is limited to 64k tokens.
Questions in a single request are answered in parallel, so dependent questions need sequential calls.
Competitor: OpenAI Decisions API (Oct 2026).
Beware of look-alike sites. A "Jev AI" site exists that is not affiliated with TypeSafe.
The name refers to the Jevons paradox: cheaper decisions should lead to more of them being automated.
## Can it run locally?
No. Jev is closed and hosted. The closest local option is an open-source zero-shot classifier:
| Model | License | Notes |
| GLiClass v3 (Knowledgator) | Apache-2.0 | All labels in one forward pass; scales to many labels |
| ModernBERT-large-zeroshot-v2.0 | see model card | NLI-style, 8k context |
| DeBERTa-v3-large-zeroshot-v2.0 | MIT | NLI-style; one pass per label, so it slows down with many labels |
| Small open LLM (Qwen, Llama) with constrained output | varies | Use per-option log-probs as scores; flexible but slower |
Key gap: none of these promise calibrated probabilities, which is Jev's main selling point. If you need calibration, fit temperature scaling on a labeled set of your own.
## Run the example (CPU is fine)
bash
python -m venv .venv
# Windows: .venv\Scripts\activate | Linux/macOS: source .venv/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt
python classify.py
classify.py answers a typed question (support-ticket routing) for a few sample texts, prints a score for each label, and reports latency.
> Research snapshot from October 2026. Jev performance and adoption figures are self-reported by TypeSafe. No independent benchmark or architecture paper existed at the time of writing.
## What Jev is
A discriminative model. It does not generate text. You give it unstructured context and typed questions (boolean, choice, ordered scale), and it returns a probability distribution with a confidence score for each.
Labels are supplied at inference time (zero-shot), so it needs no task-specific training.
Reportedly trained only on synthetic data with an unpublished method (RLCD) that targets calibrated probabilities. TypeSafe says it is transformer-based. Weights are closed.
Released 15 Sep 2026 as a hosted API only. There is no self-hosting.
Listed pricing: US$0.042 per 1M input tokens, output free. Claimed latency is 70–500 ms. Input is limited to 64k tokens.
Questions in a single request are answered in parallel, so dependent questions need sequential calls.
Competitor: OpenAI Decisions API (Oct 2026).
Beware of look-alike sites. A "Jev AI" site exists that is not affiliated with TypeSafe.
The name refers to the Jevons paradox: cheaper decisions should lead to more of them being automated.
## Can it run locally?
No. Jev is closed and hosted. The closest local option is an open-source zero-shot classifier:
| Model | License | Notes |
| GLiClass v3 (Knowledgator) | Apache-2.0 | All labels in one forward pass; scales to many labels |
| ModernBERT-large-zeroshot-v2.0 | see model card | NLI-style, 8k context |
| DeBERTa-v3-large-zeroshot-v2.0 | MIT | NLI-style; one pass per label, so it slows down with many labels |
| Small open LLM (Qwen, Llama) with constrained output | varies | Use per-option log-probs as scores; flexible but slower |
Key gap: none of these promise calibrated probabilities, which is Jev's main selling point. If you need calibration, fit temperature scaling on a labeled set of your own.
## Run the example (CPU is fine)
bash
python -m venv .venv
# Windows: .venv\Scripts\activate | Linux/macOS: source .venv/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt
python classify.py
classify.py answers a typed question (support-ticket routing) for a few sample texts, prints a score for each label, and reports latency.
收录徽章（可选）
想把收录标记放进 README 就复制这行，可选，不影响收录。
[](https://anclum.com/zh/s/vnmartinez-jev-local-alternatives/)
- 来源
- github · https://github.com/vnmartinez/jev-local-alternatives
- 收录于
- 今天