Google releases Gemma 4, a family of open models built off of Gemini 3
When Google released Gemini 3 Pro at the end of last year, it was a significant step forward for the company's proprietary large language models. Now, the company is bringing some of the same technology and research that made those models possible to the open source community with the release of its new family of Gemma 4 open-weight models. Google is offering four different versions of Gemma 4, differentiated by the number of parameters on offer. For edge devices, including smartphones, the company has the 2-billion and 4-billion "Effective" models. For more powerful machines, there’s the 26-billion "Mixture of Experts" and 31-billion "Dense" systems. For the unfamiliar, parameters are the settings a large language model can tweak to generate an output. Typically, models with more parameters will deliver better answers than ones with less, but running them also requires more powerful hardware. With Gemma 4, Google claims it's managed to engineer systems with "an unprecedented level of intelligence-per-parameter." To back up this claim, the company points to the performance of Gemma 4's 31-billion and 26-billion variants, which claimed the third and sixth spots respectively on Arena AI's text leaderboard , beating out models 20 times their size. All of the models can process video and images, making them ideal for tasks like optical character recognition. The two smaller models are also capable of processing audio inputs and understanding speech. Separately, Google says the Gemma 4 family is capable of generating offline code, meaning you could use them to do vibe coding
Google has recently unveiled its new family of open-source models, Gemma 4, built upon the foundation of its proprietary Gemini 3 Pro models. This release marks a significant shift in the company's approach to AI, as it brings cutting-edge technology and research to the open-source community. The Gemma 4 family consists of four different versions, each offering a unique set of capabilities tailored to various hardware configurations.
The Gemma 4 models are distinguished by the number of parameters they contain. Parameters are essentially the settings a large language model can adjust to generate an output. Generally, models with more parameters tend to provide better results, but they also require more powerful hardware to run efficiently. Google claims that Gemma 4 models have achieved an "unprecedented level of intelligence-per-parameter," meaning they deliver exceptional performance even with fewer parameters compared to other models in the market.
To validate this claim, Google cites the performance of its 31-billion and 26-billion Gemma 4 variants on Arena AI's text leaderboard. These models secured the third and sixth positions, respectively, outperforming models that are 20 times larger in size. This impressive achievement underscores the efficiency and effectiveness of the Gemma 4 architecture.
The Gemma 4 family is designed to be versatile and adaptable to a wide range of applications. All models can process video and images, making them suitable for tasks such as optical character recognition (OCR). Additionally, the two smaller models, with 2 billion and 4 billion parameters, are capable of processing audio inputs and understanding speech. This multimodal capability opens up new possibilities for applications that require interaction through text, image, and audio.
One of the standout features of the Gemma 4 models is their ability to generate offline code. This means users can utilize the models for tasks like "vibe coding" without requiring an internet connection. This offline functionality is particularly valuable for developers and users who need to work in environments where connectivity is limited or unreliable.
Furthermore, Google has trained the Gemma 4 models in more than 140 languages, expanding their applicability to a global audience. This linguistic versatility ensures that the models can be used in diverse regions and contexts, bridging communication barriers and enabling more inclusive AI applications.
Google is releasing the Gemma 4 family under the Apache 2.0 license, a more open and flexible license compared to its previous Gemma models, which were available under a proprietary Gemma license. This change in licensing terms grants users greater freedom to modify, distribute, and commercially utilize the models, fostering collaboration and innovation within the AI community.
In conclusion, Google's Gemma 4 family represents a groundbreaking step in the company's commitment to open-source AI. By offering a range of models with exceptional performance, multimodal capabilities, and linguistic versatility, Gemma 4 is poised to become a cornerstone of AI development for both individual developers and enterprises. The move to an open-source license further solidifies Google's dedication to promoting collaboration and democratizing access to advanced AI technologies. As the Gemma 4 models gain traction, they are likely to reshape the landscape of AI applications, driving innovation and enabling new use cases across industries and regions.
