THX-02 Uses a Gate to Think Only When Reasoning Changes a 4.5B Model's Answer

The DEV Community article describes THX-02, a 4.5B decision model released under Apache-2.0 with a free API. Its authors found that always-on chain-of-thought reasoning can degrade already-correct answers, so they built a gate that decides whether deliberation is worthwhile. The gate runs inside the forward pass and fires only when the direct answer is uncertain and the question structurally requires step-by-step work. When triggered, a short thought of up to 96 tokens is generated and fed back into the model to produce a second answer distribution. The reported results show gains of about 25 points on math word problems and 4 to 6 points on other reasoning tasks, with median latency near 80 ms.