News
How we build solutions on budget computers
We study how much useful, reliable work a local AI model can do on an inexpensive computer, and compare it with a machine where the whole model fits in memory.
Three computers take part. A budget PC with an RTX 3060 graphics card runs under tight memory limits. A second computer with the Hermes agent analyses the results, verifies the work and makes improvements. A Mac Studio M3 runs the same tasks in parallel as a reference with ample memory. All machines use the same model files and the same Strata version.
What matters to us is not speed for its own sake, but the quality and reliability of answers on real-world tasks. Preliminary results show that limited hardware does not necessarily mean lower answer quality.
A separate part of the experiment looks at whether the agent can find its own errors, fix them and prevent similar ones. Without being asked, it also began to account for the tokens its improvements cost and to cut unnecessary use.
These experiments lead to real results and practical scenarios, for example work with documents.