Jev, the inference pattern that has dominated AI Twitter for months, reduces to a straightforward operation: load a model, construct a prompt with labeled choices, extract token logits for those choices, and convert them to probabilities. A blog post by Duarte O. Carmo at NobodyWho demonstrates the full stack in 25 lines of Python, using Qwen3-0.6B running locally via llama-cpp-python.
The implementation loads a quantized GGUF model from Hugging Face, tokenizes a prompt that presents the model with discrete options, and then extracts raw logits from the model's final token. The critical step is isolating logits for only the tokens corresponding to the choice labels—in the example, the single characters "A", "B", and "C"—rather than computing probabilities across the entire vocabulary. Those logits are then normalized using log-sum-exp to produce calibrated probabilities.
The example classifies an email as legitimate, spam, or phishing. After the model processes the prompt, the code retrieves logits for the three label tokens, converts them to log-probabilities, and exponentiates to get final probabilities. The output shows the model assigned 0.885 probability to phishing, 0.084 to spam, and 0.031 to legitimate. The intermediate values—raw logits of 26.254, 27.262, and 29.614 for the three classes—are also printed, showing the unnormalized scores before probability conversion.
The stack requires Python 3.12 or later, huggingface-hub, llama-cpp-python, and numpy. The model runs with a context window of 512 tokens and logits_all set to true, which forces the inference engine to compute logits for every token position, not just the final one. The post does not report latency, throughput, or memory footprint for the Qwen3-0.6B model on any hardware, so operational costs remain unspecified.
The author frames this as a parody, noting that the post is a simplified take on Jev and directing readers to three more complete open implementations: OpenJev, openjev-sglang, and OpenJev on DiffusionGemma. The NobodyWho post does not claim to be the authoritative or production-ready version, and it does not discuss how to handle cases where the model's top token for a choice label is not the label itself, or how to scale this pattern to larger choice sets or longer option descriptions.
The takeaway for architects is that choice classification via logit extraction is implementable in minimal code and runs entirely locally, but production deployments need to measure latency and memory on target hardware, handle edge cases where token alignment fails, and decide whether the probabilities from a small model are calibrated enough for the downstream task.