Gemma 4 audio hits iOS GPU wall

This looks less like a Gemma 4 limitation and more like a mobile runtime gap: the model supports audio, but the iOS accelerator path clearly is not equally mature across frameworks yet.

• –Google’s launch materials say Gemma 4 E2B and E4B support native audio input, so the feature exists at the model level
• –CPU success plus GPU/NPU init failure usually means unsupported ops, delegate issues, or an incomplete multimodal pipeline in the runtime
• –For iOS developers, the practical takeaway is to treat CPU fallback as the baseline until the specific engine is verified on device
• –The post is a good reminder that “runs on phone” and “runs on phone GPU/NPU” are very different claims in local AI
• –If this is reproducible, the fix likely belongs in the inference stack, not in the Gemma 4 weights themselves