Qwen3 4B Outperforms 0.6B on Document QA Accuracy, but the Smaller Model Runs Faster on 6 GB GPU

A developer testing local retrieval-augmented generation on a laptop with a 6 GiB RTX 3060 compared two Qwen3 models run through Ollama. The 0.6B model generated text roughly three times faster than the 4B model, while using about a third of the peak GPU memory. On six synthetic document questions, the 4B model answered all correctly, whereas the 0.6B model passed four. The smaller model's failures included combining conditions incorrectly and returning malformed output for a case with missing information. The author notes that generation speed figures include reasoning tokens and do not reflect how quickly complete answers appear.