[Feature Request]: Gemma 4 12B

Supporting this model could be great for 12GB memory on mobile devices. MNN has been reported as the fastest decode backend compared with ik_lcpp / lcpp (it might be with OpenCL? enabled) This has a unified multimodal architecture, the model can process raw audio without an encoder module. It might be ~25 token for one second of audio, so the prefill accelerated on gpu/npu backend would be valuable for audio. There is a QAT version in the "Q4_0" weight format of lcpp which would be the best to target if possible for perceptual fidelity.