Google DeepMind Releases Gemma 4 12B, a Unified Encoder-Free Multimodal Model

Google DeepMind 发布 Gemma 4 12B：统一的无编码器多模态模型

Google DeepMind released Gemma 4 12B, a multimodal model with a unified encoder-free architecture, native audio input, Apache 2.0 licensing, and local laptop runtime with 16GB of VRAM or unified memory.

Why it matters: HKR-H/K/R all pass: the hook is local multimodal audio in 16GB VRAM, and the new architecture is concrete. It is a strong Google DeepMind open-model release, but not a frontier-model launch, so it stays below p1.