Google Releases Gemma 4 12B for Local Multimodal AI on Laptops

Google has released Gemma 4 12B, a dense multimodal model with a unified encoder-free architecture, according to a post on the Google Developers Blog. The model integrates text, vision, and audio processing directly into its large language model backbone, removing the need for separate encoders and reducing latency.

Gemma 4 12B is the first medium-sized Gemma model to support native audio input. It is designed to run locally on laptops with 16GB of VRAM or unified memory. The model is available under the Apache 2.0 license, and its weights can be downloaded from Hugging Face and Kaggle. A dedicated multi-token prediction variant has been introduced to improve local inference performance.

The release also introduces MacOS desktop applications that allow local spoken and visual interactions on Apple Silicon devices. Developers can run Gemma 4 12B as a local API server through the new LiteRT-LM tool, which supports OpenAI-compatible integrations and local execution.

Gemma 4 12B supports tasks such as speech recognition, video understanding, and coding. It can be fine-tuned using tools like Hugging Face, MLX, SGLang, and Unsloth.