I want to try a hybrid setup of Gemma 4 E4B with lots of context for general, then Qwen 3.5 9B or larger for coding. Strix Halo set up this weekend, which may enable even larger Qwen models with tons of context.

You can run it, however those low quantized models (iQ2, iQ4, Q2) will very likely underperform the 9B versions at Q6/Q8.

The larger Gemma models are quite good at PHP. I would not be surprised if that was a training objective — it's one of the more consumer-focussed programming languages. They have very good knowledge of wordpress hooks.