Google has released Gemma 4 12B, a mid-sized multimodal model designed to run locally on consumer laptops with 16GB of VRAM. Its standout feature is an encoder-free architecture that processes vision and audio inputs directly through the LLM backbone — replacing traditional separate encoders with lightweight embedding modules. The model bridges the gap between the smaller E4B and the larger 26B MoE model, offering near-26B benchmark performance at under half the memory footprint. It supports native audio input (a first for mid-sized Gemma models), Multi-Token Prediction drafters for lower latency, and is released under Apache 2.0. Developers can run it via Ollama, LM Studio, llama.cpp, MLX, vLLM, or deploy on Google Cloud. Gemma 4 models have now crossed 150 million downloads.
