Google DeepMind: Gemma 4 12B, Compact Multimodal AI

Google Deepmind recently introduced Gemma 4 12B, an open-source artificial intelligence model that promises to transform the use of AI on everyday devices. This model stands out for its ability to operate efficiently on laptops equipped with just 16 GB of RAM. By natively integrating the processing of text, images, and audio, Gemma 4 12B eliminates the need for separate encoders, significantly reducing processing time, memory usage, and latency, according to Google.

Despite its more modest size, Gemma 4 12B delivers performance comparable to that of the 26B model, which is twice its size, across various benchmarks. It is also the first mid-sized model in the Gemma series to incorporate native audio processing, a significant advancement for developers.

Among its features, Gemma 4 12B supports speech recognition, code generation, and video analysis. During a demonstration, the model analyzed a five-minute excerpt from the Google I/O conference, processing 313 images at one per second, in addition to the audio.

On benchmarks such as GPQA Diamond, MMLU Pro, and DocVQA, Gemma 4 12B approaches the performance of the 26B model while surpassing the older Gemma 3 27B model. The model operates locally and is accessible on platforms like Hugging Face, Ollama, LM Studio, and others, under the Apache 2.0 license, allowing for commercial use.