An independent developer compared Nemotron 3.5 Bolt E4B with its Gemma 4 E4B base on 299 Korean document questions using the same retrieved passages and instructions. With thinking enabled, accuracy was 67.2% and 69.6%, respectively; the difference between models was not statistically detectable. Thinking improved accuracy by roughly four percentage points while increasing time per question three- to fourfold. This single dataset and automated grading do not prove equivalence, and gains on an English-centered official benchmark cannot be assumed to transfer directly to Korean.