A blog post from NobodyWho breaks down Jev—a technique receiving attention in AI circles—by demonstrating a working implementation in just 25 lines of Python.
Jev functions as a classification system that takes a prompt with multiple choice options and outputs probability scores for each category. According to the post, the core workflow involves three steps: loading a language model from Hugging Face, preparing a prompt with labeled choices, and converting the model’s output logits into probabilities.
The example uses Qwen3, a 0.6-billion-parameter model in GGUF format, to classify an email as “Legitimate,” “Spam,” or “Phishing.” When prompted with a scenario about payroll requesting credentials on a non-company page, the model assigned probabilities of 0.031, 0.084, and 0.885 respectively—correctly identifying it as phishing with highest confidence.
The post emphasizes three practical advantages: the model runs locally on a user’s machine, processing is fast compared to API calls, and user data never leaves the device. This contrasts with approaches the authors note they did not use, such as calling external APIs, generating synthetic training data, or applying reinforcement learning techniques to calibrate outputs.
The authors acknowledge the implementation differs from more elaborate implementations elsewhere, noting this is a simplified version. They reference alternative open-source implementations like OpenJev and openjev-sglang for more complete approaches. The post carries a disclaimer that it is parody, suggesting the hype around Jev may exceed its technical novelty.
The technique appears suitable for scenarios requiring on-device text classification with explicit choice sets—useful for privacy-sensitive applications where data handling and latency matter.
Key facts
- Jev classifies text into predefined categories by computing probability scores from language model outputs
- The example implementation uses a 0.6B-parameter Qwen3 model running locally without API calls
- The technique involves tokenizing choice labels, extracting logits, and normalizing them into probabilities
- Data processing occurs entirely on the user’s device, with no external data transmission
- The post is labeled as parody, referencing more complete open-source implementations elsewhere