How well does this work with local models like https://github.com/receptron/laya (open source Jev) and vision capable models like Qwen 3.8 Flash Next, or even Qwen 3.8 27B on xhigh thinking?