Ive tested all the models where it can fit into my 4080 vram. Even some slightly bigger. Gemma4 outperforms all of them so that's what I'm sticking with for now. Gemma4 12B model judging network configs for CVE false positives stuck at ~77.8% pass rate and it's a reasoning issue. Have the complete CVE list for the code base and the device config. I pull each networks device's running config, batch ~10 CVEs per call to a local gemma4:12b (Ollama), and have the model return applicable / not_appl...