Simple Benchmark Review: Ollama on Jetson Nano

Short summary

A developer benchmarks several Ollama models (qwen2.5:3b, llama3.2:3b, mistral:7b) at varying quantizations on a Jetson Nano for quiz-generation tasks. Qwen2.5:3b-instruct achieved 100% accuracy across all tested quantizations, while llama3.2:3b peaked at 90% and mistral:7b hit 100% at q4_K_M. The author created a swap file to handle RAM limitations and notes that many model/quant combinations couldn't fit on the Nano's GPU.

• •Qwen2.5:3b-instruct scored 100% accuracy on a 10-question quiz benchmark across all quantizations on Jetson Nano
• •Mistral:7b at q4_K_M also hit 100%, while llama3.2:3b maxed at 90%
• •Swap file needed due to Nano RAM constraints; several model/quant combos couldn't fit on GPU

Generated with AI, which can make mistakes.