hearim(헤아림): Maybe you don’t need a special model for Jev — ordinary local LLMs already have the capability

What if you don’t need a new model to get Jev-style decisions? I built hearim(헤아림. hea-a-rim. not Here, I am.. lol) , an open-source gateway that turns ordinary local or hosted LLMs into a Jev-compatible decision API. Give it some data and a set of choices, and it returns a decision with probabilities—using models you can already run through llama.cpp, vLLM, SGLang, Ollama, or an OpenAI-compatible server. No task-specific fine-tuning model. No custom decision-model weights. No parsing generated JSON and hoping the model followed the schema. only pre-fill (with prefix caching) + only 1 token decode (with logit) This kill most time-latency and no need to generate text. The basic idea is that a pretrained language model already learned a great deal before instruction tuning or RLHF taught it how to chat. If you only need it to choose between clearly defined alternatives, much of that capability may already be present in its probability distribution. The model may already know. We just need to stop asking it to write an essay. hearim turns a decision into a constrained set of alternatives, reads the model’s probabilities for them, and returns the result through the Jev Choice , Score , and Noul API format. I’ve been using this underlying technique for more than a year in a separate implementation —before Jev was announced. hearim itself is a new, independent reimplementation , built specifically around the Jev API contract. I’m not claiming that it reproduces Jev’s architecture, training, calibration, speed, or performance. The interesting claim is narrower: A useful Jev-style interface does not necessarily require a specialized Jev-style model. The implementation, supported backends, limitations, benchmarks, and technical details are here: GitHub: github.com/ziozzang/hearim Paper: github.com/ziozzang/hearim/blob/main/paper/paper.pdf I’d love to see what happens when people try it with different local models—especially smaller base models, not only instruction-tuned ones. The model may already know. We just need to stop asking it to write an essay.