Google claimed Gemma 3 outperformed Meta’s Llama-405B, OpenAI’s o3-mini, and DeepSeek-V3 in preparatory human most like evaluations on LMArena.
Google has launched Gemma 3, the present day addition to its family of light-weight open models, which are designed to run right on phones, laptops, and different devices.
The Gemma 3 collection of models are powered by using the identical research and technology that powers Google’s Gemini 2.0 models the tech large stated in a blog post published on Wednesday, March 12. “This supports you to generate interesting consumer experiences that could fit on a single GPU or TPU host,” Google stated.
Gemma 3 models are accomplished of processing text content and visual inputs but can only generate text outputs. The models are to be had in one billion, four billion, 12 billion, and 27 billion-parameter sizes, permitting developers to pick the model quality desirable to run their AI programs.
“The 27B model was trained with 14 trillion tokens, the 12B model was trained with 12 trillion tokens, 4B version was trained with 4 trillion tokens and 1B with 2 trillion tokens,” in line with the Gemma 3 page on Hugging Face.
While Gemma 3 has been trained on a dataset of text statistics, Google did no longer specify the data sources. It additionally disclosed that the weights of Gemma 3 are open-source and may be utilized by developers to generate pre-trained variations and instruction-tuned variations of the small language model (SLM).
Google similarly claimed that Gemma 3 comes with a 128k-token context window that lets model to understand larger amounts of data.
How Gemma 3 compares with different AI models
Google has claimed that Gemma 3 attained a higher benchmark score than Meta’s Llama-405B model as well as OpenAI’s o3-mini and DeepSeek-V3 in preliminary human preference evaluations on LMArena, an open platform evolved via UC Berkley researchers for crowdsourced AI benchmarking.
Gemma 3 can be used to generate AI programs that examine images, textual content, and short videos. It can also manage the linguistic tasks as Gemma 3 helps over 35 languages, with pretrained support for over 140 languages. In addition, developers can use Gemma 3 to generate AI equipment that automate tasks and provide AI agent-based abilities due to its based outputs and characteristic calling support.
Gemma 3 is available for download via platforms such as Kaggle and Hugging Face. It is also handy by Google Studio.
“Gemma 3 gives a couple of deployment options, along with Vertex AI, Cloud Run, the Google GenAI API, Local environments and other platforms, giving you the flexibility to pick out the great fit for your program and infrastructure,” the enterprise stated.
The series of models may be further trained and exceptional-tuned the use of platforms such as Google Collab, Vertex AI, and even on gaming GPUs, as in step with the enterprise. “Gemma 3 ships with a made over codebase that consists of recipes for efficient fine-tuning and inference,” it introduced.
What is ShieldGemma 2?
Alongside the launch of Gemma 3, Google introduced the rollout of its 4 billion parameter-size AI safety device referred to as ShieldGemma 2.
ShieldGemma 2 can be utilized to attach labels consisting of dangerous content, sexually specific, and violence to AI-generated images. Google claimed that ShieldGemma 2 may be incorporated with some other tool used by developers, with options for further customization available.
