LocalAI Qwen 3.5 35B benchmark: Vulkan wins

This is less a universal backend verdict than a workload split, and that matters. If your app is chatty and decode-heavy, Vulkan looks like the better default; if your pipeline is prompt-heavy, ROCm still has an edge.

• –The result held across two different Qwen3.5-35B variants, so the backend pattern looks real rather than quant-specific noise.
• –Vulkan’s roughly 10-15% generation lead is the number to watch, because token streaming is what users feel in interactive sessions.
• –ROCm’s prompt-processing advantage is still meaningful for ingestion, long-context preprocessing, and batch-style local workflows.
• –Prefix caching and 200K-context tests make this relevant for agentic use cases, not just short chat prompts.
• –For Strix Halo and similar AMD APUs, backend choice should now be workload-specific instead of assumed.