I’ve been doing lots of testing back and forth with this 7900xtx. All of my workloads were relying on qwen3.6 models, which are amazing fwiw, but I wanted some diversity in thought. Namely for Honcho workload tiers and differing cron jobs. Not every workload benefits from an agentic-tuned model, so I’ve been testing out Gemma 4 models more. They also dropped quantization-aware training versions of the Gemma 4 family, which reportedly maintain the fidelity of BF16 weights, but with Q4 weights. ...