← All topics

#performance

  1. Measuring why local AI felt slow

    A slow greeting pushed me to separate model loading, instruction processing and queued work from response generation.