My weather lookup took 224 ms. My local AI answer took 26 minutes.
By William Paul “Bill” Herald
I have been testing what my own computer can do with AI. A weather report seemed like a useful place to start: get an observation, then give me an answer I could read. I asked for the observed weather in Bloomsburg, Pennsylvania, using GLM-5.2 through Colibri. The completed request used Bloomsburg's coordinates. A weather tool returned a National Weather Service observation as JSON, and the local model turned that data into a short answer with the station, conditions, time and source.
The weather lookup took 224 milliseconds. The whole answer took about 26 minutes. A person using the system waits for the answer, not just the lookup.
For a weather report, 26 minutes is a long wait. The observation was already about an hour old when the tool fetched it. By the time the answer arrived, I would have checked again. This answer passed the check I set, but earlier attempts failed or timed out. Model startup was outside the timed run. I need more completed runs before I can say what someone should normally expect.
I have spent more than 40 years building software. A fast database query could still leave someone waiting for a slow page. I want to test local AI by the same practical measure: Did it finish the job, and how long did someone wait?
A separate test made an immediate repeat look fast. The first try took almost eight minutes; the repeat took about seven seconds. Both were limited to one output token, so neither was a finished answer. The repeat reused a prompt prefix and caches I had not cleared. A CPU-only check also answered one readiness prompt in about 22 seconds, then missed a five-minute deadline on a different prompt. Those were useful clues, but not a fair comparison of setups.
Next I want to give each setup the same useful task and deadline, including a smaller model. I would count startup, failed attempts and corrections, then measure electricity used by the whole computer before making an energy claim.
AI helped me organize the test and inspect the logs. The timings came from the recorded run; I checked what the records support.
I want local AI to keep more of my personal information on my own computer. I also need it to answer while the information is still useful. This test showed me where to focus next.
My measurement notes give the exact figures and limits. My broader article on local AI capacity explains why I keep working on this.