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.