Ollama on Jetson Nano: qwen2.5:3b-instruct tops small benchmark for quiz generation

A developer ran informal benchmarks of multiple large language models via Ollama on a Jetson Nano to find the best fit for a local, free flashcard and quiz-generation app. Due to limited RAM on the device, a swap file was created as a safety net before testing could begin. Models including qwen2.5, llama3.2, and mistral variants were tested across different quantization levels using a 10-question OSI model quiz as the reference. qwen2.5:3b-instruct achieved 100% accuracy across all tested quantizations, while some larger models could not run at all due to GPU memory constraints. The developer plans to switch their app from llama3.2:3b-instruct to qwen2.5:3b-instruct based on these results, acknowledging the test covers only one narrow use case.