I'm a real estate investor, and I use AI to acquire properties and run my business. I currently use 3 x $20 subscriptions. Anthropic, Open AI and Nous Portal. Eventually I will use just Nous Portal @ $20 a month, but it won't fully cut it for my needs. So I set out to repurpose some old hardware. - Hardware / Software specs at the bottom. I slapped a couple of Nvidia 3070's w/8G ea. into an old gaming system I had laying around. I was worried about having the 16G split across two 8 GB cards over the PCIe lanes, but both buses are 16x and the CPU supports 40 PCIe 3.0 lanes. "Oldie but a goodie" I had to use Ubuntu 24.04 since 26.04 had all kinds of issues with this dinosaur. I finally got the operating system on the box, hardware tested and verified and got it stabilized with no overclocking, all base stuff. The best model I could find that fit on my old hardware was the Gemma 4-12B. It runs 100% Vram with zero CPU bleed. So I was ready to run my use cases against it versus the cloud. I ran Two side-by-side tests on my Web Scraping of Zip codes to find properties. Tested 85053 and 85022 using one live, read-only collection. The resulting facts were frozen and sent identically to: Old method: GPT-5.6 New method: Nexus gemma4:12b This avoids listing changes contaminating the comparison. Live scan outcome 11 listings discovered 11 met dimensional criteria 7 verified no-HOA qualifiers 4 excluded for HOA No database writes No email sent No paid Firecrawl No runner remains active Model comparison Semantic accuracy • Gemma 4 12B: 100% • GPT-5.6: 100% Correct top-three ranking • Gemma 4 12B: Yes • GPT-5.6: Yes Detected report-count discrepancy • Gemma 4 12B: Yes • GPT-5.6: Yes Correctly identified dry-run/no email • Gemma 4 12B: Yes • GPT-5.6: Yes Unsafe claims • Gemma 4 12B: None • GPT-5.6: None Exact machine contract • Gemma 4 12B: 100% • GPT-5.6: 74% Average latency • Gemma 4 12B: 17.7 sec • GPT-5.6: 25.7 sec Gemma was approximately 8 seconds—or 31%—faster.