Burl
Menu

Local AI on a 2018 workstation

A 2018 workstation, a firmware fix, and an evaluation that cancelled a purchase. Recall went from 17% to 83% after one prompt rule.

Local AI on a 2018 workstation, the firmware fix, the 24 GB card, and the measured recall and speed changes

We were sizing a local AI box. The real problem was one rule missing from a prompt.

  • The shadow test ran four models on real traffic. The card that went silent on a quarter of the hard cases was dropped from the plan.
  • The prompt fix took recall from 17% to 83%, in ten minutes.
  • The 2018 Dell Precision workstation would not see its 24 GB card at all. The BIOS gave the card too small a window of memory. We found the hidden firmware value that fixes it, and published the fix.
  • Tuning moved one model from 49 to 62.5 tokens a second on the same hardware. (Our own box, October 2026.)

The full story is on the Local AI page.

Want something like this?

A sentence is enough. We'll reply within a business day, and if we're not the right people, we'll say so.

Get in touch