Google Introduces Gemma 4 Open Models to Compete with Larger AI Systems

Daily Technology

03/04/2026

Google has announced the release of Gemma 4, a new family of open-weight artificial intelligence models derived from the same research and technology that powers its proprietary Gemini models. This launch introduces a series of more efficient systems designed to offer high performance within a smaller computational footprint, making advanced AI more accessible to the open-source community.

Performance and Model Variants

The Gemma 4 family consists of four distinct models tailored for different hardware capabilities. For edge devices like smartphones, Google offers the 2-billion and 4-billion parameter "Effective" models. For more demanding tasks on powerful hardware, the lineup includes a 26-billion parameter "Mixture of Experts" model and a 31-billion parameter "Dense" system. Parameters in a large language model are the variables the model adjusts to refine its output; typically, a higher parameter count correlates with more sophisticated responses but also requires greater processing power.

Google asserts that Gemma 4 achieves an exceptional level of "intelligence-per-parameter." As evidence, the company highlights the performance of its 31-billion and 26-billion parameter variants on the Arena AI text leaderboard. These models secured the third and sixth positions, respectively, outperforming other models that are reportedly up to 20 times larger in size. This benchmark data suggests a significant advancement in optimizing AI model efficiency without compromising output quality.

Advanced Capabilities and Licensing

Beyond text generation, all Gemma 4 models are equipped with multi-modal capabilities, including the ability to process video and images for tasks such as optical character recognition (OCR). The two smaller models can also process audio inputs and understand speech. Furthermore, the entire family is capable of generating offline code, enabling development tasks without an internet connection, and has been trained in over 140 languages.

A notable change with this release is the adoption of the Apache 2.0 license, a shift from the custom Gemma license used for previous versions. This move grants developers significantly more freedom and flexibility. According to Google, the Apache 2.0 license allows for complete control over data, infrastructure, and models, enabling developers to build and deploy their applications securely in any environment, from on-premises servers to the cloud. The model weights for the Gemma 4 family are now available through platforms such as Hugging Face, Kaggle, and Ollama.

2026-03-24

New Five-Level Framework Revolutionizes Humanoid Robot Evaluation

A new five-level model is introduced to evaluate humanoid robots across mobility, manipulation, and cognition, setting new benchmarks for the industry.