Question 1

What is OpenJev in one sentence?

Accepted Answer

An open decision engine: point it at a small local model or any OpenAI-compatible API you already have a key for, send a state plus typed questions, and get type-safe answers with raw softmax probabilities. It also ships a crush radar for WeChat, QQ and Telegram chats that returns a verdict plus next-move advice.

Question 2

How is it different from the official Jev cloud?

Accepted Answer

Same interface, different engine. The README's table: OpenJev uses masked-logit softmax over any open LLM, runs 100% locally at $0 with roughly 0.1–3 s latency, while the closed cloud reports 70–500 ms, $0.042 per million input tokens and 64K context against Qwen3's 32K. A RemoteJev client still calls the official cloud.

Question 3

Can it hallucinate or return malformed answers?

Accepted Answer

The README says hallucination is impossible by design because OpenJev never decodes: it reads the raw logits at the first answer position, masks the answer-token logits and softmaxes them. Probabilities are raw softmax values, not sampled, prompted for, or parsed from text, and type errors are described as mathematically impossible.

Question 4

What hardware or key do I need?

Accepted Answer

The default local brain runs any Hugging Face causal LM or local path; the README says a 0.6B model needs about 3 GB RAM and shows a CPU-only fp32 demo with Qwen3-0.6B needing no key. The OpenAI-compatible route takes any key or local server, and the official cloud route takes a TYPESAFE_API_KEY.

Question 5

What does OpenJev cost to run?

Accepted Answer

Local inference is $0 and nothing leaves the machine. API-backed runs cost per token according to the provider — OpenAI, OpenRouter, DeepSeek, Groq, Together, Ollama, LM Studio, vLLM or llama.cpp — and self-hosted servers stay local. The official-cloud route is billed at Jev's pricing, which the README lists at $0.042 per million input tokens.

Question 6

How do I try OpenJev quickly?

Accepted Answer

Python 3.9+ and the README's quick start. A captured run is python demo.py, printing department, frustration and urgency answers with confidences, plus forward passes, tokens and latency. The three backends share one .system_one(state, questions) call and a POST interface; see the repository README and its Hugging Face Space for full setup.

Question 7

What is the crush radar feature?

Accepted Answer

A use case bundled with the engine: it reads WeChat, QQ or Telegram chats and returns a verdict on the situation plus next-move advice. The GitHub description frames it as a pursuing radar with verdict calls like charge, hold, slow or stop, and a Hugging Face Space is linked as the demo.

OpenJev — lookski

Highlights

Quickstart

Questions about this entry

From the Jev guides

From the Jev AI model community