Everyone is talking about Jev. Jev this, Jev that. Everyone on Twitter is all over Jev, how it’s the next frontier of large language models and the AI paradigm. So a developer named Duarte O.Carmo decided to ask: what exactly is Jev? His answer arrives as a parody blog post called “Jev in 25 Lines of Python.”
The post loads a language model, feeds it a prompt, and asks it to pick a choice. That’s it. The whole thing.
Model Setup
The code uses the Llama library and a GGUF model from Qwen, specifically the Qwen3-0.6B model. Duarte loads the model with a few parameters: context length set to 512, logits all enabled, and no verbose output.
Prompt Construction
The prompt asks the model to pick one of three options. The example email is a generic warning about a payroll sign-in page, but the model never sees it directly — the prompt is constructed from the text and the choices.
Choice Output
The model computes logits for each label, converts them to probabilities, and ranks the options. The result: one option that looks likely, one that looks medium, and one that looks unlikely.
Verdict
Duarte’s point is direct. The model classifies. It’s fast. It’s local. You don’t send your data anywhere else.
“Yeah, we know.”
What Jev Isn’t
Duarte’s list of absences is as telling as the code itself. He does not call the model a System One decision model. He does not call an API. He does not create synthetic data. And he does not train a model with Reinforcement Learning for Calibrated Decisions (RLCD).
The post is a parody, and the author clearly enjoys making the point through the code rather than stating it outright.
Comparison Table
| Feature | Jev in 25 Lines of Python |
|---|---|
| Model Type | Existing language model like Llama |
| Training | None shown |
| Calibration | Not claimed |
| API | None called |
| Data Sent | None |
Open Source Angle
Duarte’s project is part of NobodyWho, an open-source project. He asks readers to leave a star on Github to support the work. He also points to several other open implementations of Jev, including OpenJev and a version on DiffusionGemma.
Verdict on Jev
The parody is effective because it is so close to the truth. A language model can produce probabilities for choices, and it can do so in a handful of lines of code. But Jev is a specific system with a specific history, and it is not just a model that happens to output probabilities.
The post does not claim Jev is useless. It claims the hype around it is overblown. It is a narrow point, but it is made with wit and a short script.
Jev in 25 Lines of Python is a fun read and a sharp one. It is a reminder that the tools we use are often simpler than the stories we tell about them. The model does the work. The code does the work. The probabilities are there. The calibrated part is not claimed.
Source material: “Jev in 25 Lines of Python,” nobodywho.ai.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.