Coin World News: Google's latest Gemma 4 technical report demonstrates how to use a minimalist architecture to make large models run faster on phones and edge devices while saving more memory. The Gemma 4 12B model achieves a "encoder-free" minimalist integration of both visual and audio modalities for the first time, directly cutting the visual translator by 93% and completely eliminating the audio translator, solving the VRAM waste caused by "external translators." In addition, the model reduces memory usage by 37.5% when processing long texts through global attention reuse. Its lightweight edge version's prediction accelerator also reduces the dictionary range from 260k to 4,096, significantly improving decoding speed.