The AI Capacity We Already Own

By William Paul “Bill” Herald

I have been testing what my own computer can do with AI. It has a processor, memory and storage I already paid for. I want to put those resources to work, keep more of my personal information under my control, and find out where the limits are.

I contribute to Colibri, a project that helps run large AI models on a computer you own. I have submitted fixes and tests for its memory planning and Windows behavior. The public contribution record shows that work. Getting a model to run is a start. I also want to know whether it can finish a useful job in time.

A weather report put it to the test

I ran GLM-5.2 through Colibri and asked for the observed weather in Bloomsburg, Pennsylvania. The completed request used the town's coordinates. A weather tool returned a National Weather Service observation as JSON. The model turned it into an answer I could read, with the station, conditions, observation time and source.

The lookup took 224 milliseconds. The answer took about 26 minutes. If I were checking the weather before going out, I would have checked again long before the answer arrived. The observation was already about an hour old when the tool fetched it.

Earlier attempts failed or timed out, and model startup was outside that timed run. One successful answer does not tell me what to expect every day. My weather article explains the experience; the measurement notes keep the exact figures and limits. AI helped me organize the test and inspect the logs. I checked the timing against the recorded run and tool responses.

What I want to learn next

Colibri can use a computer's CPU, memory and storage to run large models. A GPU helps on some paths, but Colibri's documentation shows ways to run without one. That makes the capacity we already own worth examining. It does not make every setup fast.

I am also interested in models such as Kimi K3, which select only part of their expert network for each piece of text. I want to see whether that design helps on ordinary hardware. I have not measured an electricity saving. Running a trained model locally also does not erase the energy used to train it, and moving model data from storage takes work. Research on storage offloading is useful background, not a measurement of my computer.

My next test will give each setup the same useful task and deadline. I will count startup, failed attempts and corrections, and compare first use with a repeat. I will include a smaller model and measure electricity for the whole computer before making an energy claim.

Local AI gives me more say over where my information goes, although connected tools can still send it elsewhere. I will keep testing what this computer can finish, how long it takes and what it costs. I will publish the results with the limits I can verify.