What we learn building an AI assistant that runs on ordinary Windows hardware — measurements, trade-offs, and the parts that did not work.
Our full engineering research report. We treat NVMe, RAM, pinned host memory and VRAM as one coordinated hierarchy so a 35B sparse model can run on a 10 GB card — then show, with measurements, why most of the aggressive caching ideas were slower than simply placing layers correctly. Includes the negative results.
Read the reportThe arithmetic behind the answer, what happens when a model nearly fits, and the throughput cliff we measured when one extra layer was pushed onto the GPU: 52 tokens per second became 5.
Read itFour things genuinely change when the model moves onto your desk, and three things do not change in your favour. Written by people selling the local option, which is exactly why the second list is here.
Read itNo desktop AI tool is at zero packets. Here is every connection Lemonade makes, why it makes it, and how to check any vendor's claim yourself in about two minutes.
Read itLemonade is a one-off $49 licence with a 14-day free trial. No datacenter, no subscription, no upload.
Also available as RSS.