Gemma 4 12B, the encoder-free multimodal that fits on a laptop

Google DeepMind releases Gemma 4 12B, an encoder-free multimodal model running locally on 16 GB VRAM to process text, vision, and native audio offline.

No visual encoder, no audio encoder: Gemma 4 12B feeds images and sounds directly into the model's backbone. Google DeepMind thus bridges the gap between the E4B designed for edge devices and the 26B Mixture of Experts, with an intermediate-sized model that is also the first in the series to natively accept audio.

Where classic multimodal models rely on separate encoders, which are sources of latency and memory consumption, vision here is handled by a simple embedding module (a matrix multiplication, positional embeddings, and normalizations) and the raw audio signal is projected directly into the text token space. The result approaches the 26B's performance on standard benchmarks for less than half the memory footprint and runs locally with 16 GB of VRAM or unified memory, with Multi-Token Prediction drafters further reducing latency.

The official demonstration shows it transcribing, formatting, and translating voice entirely offline. Released under an Apache 2.0 license, the model can be downloaded from Hugging Face and Kaggle, runs via LM Studio, Ollama, llama.cpp, MLX, or vLLM, can be fine-tuned with Unsloth, and comes with an official Skills Repository for agents, with the Gemma 4 family having also just surpassed 150 million downloads.