Gemma 4 12B ditches multimodal encoders, runs locally on 16GB laptops

Google released Gemma 4 12B, a mid-sized open-weight model that drops the conventional approach of using separate vision and audio encoders. Instead, image and audio inputs feed directly into the LLM backbone — vision goes through a lightweight embedding module built from a single matrix multiplication plus positional embeddings, while raw audio is projected straight into the text token space. Google argues this trims latency and memory overhead compared to traditional multimodal pipelines.

The model sits between the edge-oriented E4B and the 26B MoE flagship, and Google claims benchmark numbers approaching the 26B variant at less than half the memory footprint. It fits on consumer laptops with 16GB of unified memory or VRAM, ships with Multi-Token Prediction drafters for faster inference, and is Apache 2.0 licensed. It is also the first mid-sized Gemma to accept native audio input.

Distribution is broad from day one: weights on Hugging Face and Kaggle, runtime support across Ollama, LM Studio, llama.cpp, MLX, vLLM, SGLang, and Unsloth, plus a Skills Repository aimed at agentic workflows. Google says the Gemma line has now passed 150 million downloads.