Gemma 4 12B is a medium-sized multimodal model developed by Google DeepMind, representing a significant architectural shift in the Gemma open-weights family. Released in June 2026, it utilizes a unified, encoder-free architecture that processes text, image, audio, and video data within a single decoder-only transformer backbone. By eliminating separate, heavy vision and audio encoders, the model achieves lower latency and a reduced memory footprint, making it suitable for local execution on consumer hardware such as laptops with 16GB of VRAM.
A core feature of the model is its advanced reasoning capability. The instruction-tuned variant supports a native "Thinking Mode," which allows the model to perform internal chain-of-thought processing before generating a final response. This enables more reliable performance on complex mathematical, scientific, and coding tasks. Additionally, the model features a 256,000-token context window, allowing for the analysis of long documents and large-scale multimodal inputs.
Architecturally, Gemma 4 12B consists of 48 layers and approximately 11.95 billion parameters. It incorporates modern transformer enhancements such as grouped query attention (GQA) and interleaved global and local attention layers to optimize memory efficiency during long-context inference. Released under the Apache 2.0 license, it is designed for a variety of agentic workflows and local-first AI applications, providing frontier-level intelligence in a developer-friendly size.