Google’s Gemini Tech Powers New Open-Weight Models

Google has released the Gemma 4 family of open-weight models, bringing technology derived from its Gemini 3 large language models to the open source community. The new models are designed to run on a range of devices, from smartphones to workstations, and offer varying levels of performance based on the number of parameters.

The Gemma 4 family consists of four models: 2-billion and 4-billion parameter “Effective” models for edge devices, and 26-billion “Mixture of Experts” and 31-billion “Dense” models for more powerful machines. Google claims the new models achieve “an unprecedented level of intelligence-per-parameter.”

According to Google, the 31-billion and 26-billion parameter Gemma 4 variants achieved third and sixth place, respectively, on Arena AI’s text leaderboard, surpassing models with 20 times more parameters. This suggests a significant leap in efficiency and performance for open-weight language models.

Multimodal Capabilities and Language Support

All Gemma 4 models can process both video and images, making them suitable for tasks like optical character recognition. The smaller 2-billion and 4-billion parameter models also support audio input and speech understanding. The models are capable of generating offline code, allowing for coding tasks without an internet connection.

Google has trained the Gemma 4 family on data spanning more than 140 languages, expanding its potential applications for global communication and content creation.

Open Source Licensing and Availability

The Gemma 4 family is released under the Apache 2.0 license, a permissive open-source license that allows for broad use and modification. This contrasts with Google’s previous Gemma models, which were released under a custom Gemma license. The move to Apache 2.0 is expected to encourage wider adoption and collaboration within the open source community.

This release follows Google’s launch of Gemini 3 Pro at the end of last year, which represented a significant advancement in its proprietary large language models. By extending some of the core technologies and research behind Gemini 3 to the open-source Gemma 4 models, Google aims to foster innovation and accessibility in the field of artificial intelligence.

Recent Developments in Google’s AI Model Family

The release of Gemma 4 comes amid a flurry of activity in Google’s AI model development. In January 2026, Google introduced MedGemma 1.5 4B, a compute-efficient model for high-dimensional medical imaging interpretation. Also in January, TranslateGemma, a collection of translation models supporting 55 languages, was unveiled.

December 2025 saw the release of Gemma Scope 2, designed to aid AI safety research, FunctionGemma for bespoke function calling, and T5Gemma 2, the next generation of encoder-decoder models. VaultGemma, described as the world’s most capable differentially private LLM, was also released in September 2025.

These releases demonstrate Google’s commitment to developing a diverse portfolio of AI models catering to a wide range of applications and research areas. The Gemma family, in particular, plays a crucial role in making advanced AI technology accessible to a broader audience through open-source licensing.

The development of Gemini, Google DeepMind’s family of multimodal large language models, began in May 2023 as a successor to PaLM 2. Gemini is unique in its ability to process multiple data types simultaneously, including text, images, audio, video, and code. The latest stable release of Gemini is version 3.1 Pro, 3 Deep Think, 3 Flash, and 3.1 Flash lite, released on March 3, 2026.