Open-Source Logit Extraction Replicates Commercial Decision AI Models Like Jev
Developer TheoLeeCJ created OpenJev, an open-source project demonstrating that extracting logit probabilities for choice options from local Qwen models can match or outperform commercial decision models like TypeSafe AI's Jev. The implementation maps decision choices to specific letters and captures the model's raw logit probabilities directly in the browser or locally. Commercial decision models like Jev charge for providing calibrated decision probabilities instead of standard text generation. Demonstrating that local open-source LLMs can achieve similar probability-based decisions through basic logit extraction enables developers to build fast, cost-effective decision infrastructure without proprietary APIs. OpenJev works by assigning each multiple-choice option a designated token (such as letters A, B, or C) and inspecting the logit probability distribution across those tokens at the critical output step. The author published the code on GitHub alongside a web-based demo using a 4B-parameter Qwen model.
## BACKGROUND
TypeSafe AI recently introduced Jev, a commercial System One model built to return fast, typed probabilities for automated decision-making rather than chat responses. In large language models, logits are the unnormalized numerical outputs produced before calculating token probabilities via a softmax function. Extracting these logits allows developers to evaluate how strongly an LLM favors pre-defined choices without waiting for full text generation.