JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Chinese implementation and benchmark guide

Guide to measuring the speed of an image sorting AI assistant

This guide explains how to measure the speed of an AI assistant that looks at an image and picks a category. It shows how to separate the software's thinking time from the total request time.

Original by RJMSWDEvaluation

Listen to this guide

JevMade’s plain-English explanation

0:00 /

AI narration

Credits

“Qwen Choice” by RJMSWD. Read the original source.

This expanded guide is an AI-narrated adaptation prepared by JevMade. It expands the source’s essential ideas, examples and caveats in JevMade’s own words and is not a word-for-word reading. The synthetic voice does not imitate the author or imply their endorsement.

Our summary

This Chinese-language guide shows how to use an AI tool on your own computer to answer a question about an image. Its example reads a picture of a customer message and chooses whether the problem concerns billing, technical support, sales, or something else.

The software reads the image and compares the allowed answers in one step, rather than writing a response word by word. On the author's computer, half the measured requests finished in about 169 milliseconds or less. That test used one specific graphics card and one image.

This helps developers measure speed on their own equipment. Repeating the same example does not show accuracy on unfamiliar pictures, and the percentages returned are not proven chances of being right. The project borrows Jev's choice-based design but uses a different AI tool called Qwen.

Key takeaways

  1. Compare allowed answers directly instead of waiting for the software to write a reply.
  2. Measure the whole request separately from the graphics card's calculation time.
  3. Repeating one image can measure speed, but does not prove accuracy on other images.

The author reports a median speed of 168.694 milliseconds using one RTX A6000 graphics card and one image. These are local measurements, not results from a hosted service.

GitHub repository · Source reviewed

Read the original guide Opens the author’s site in a new tab.