Google Releases Gemma 3 Family of Open-Source AI Models, Can Run on a Single GPU

Google launched the Gemma 3 household of synthetic intelligence (AI) fashions on Wednesday. Successor to the Gemma 2 sequence, which was launched in August 2024, the brand new open-source fashions arrive with textual content and visible reasoning capabilities. The Mountain View-based tech big mentioned these fashions provide assist for greater than 35 languages and could be fine-tuned to supply extra assist for 140 languages. Notably, the tech big claimed that these fashions are optimised to run on a single GPU or Google’s Tensor processing unit (TPU).

Google Releases Gemma 3 Series AI Models

In a blog post, the tech big detailed the brand new small language fashions (SLMs), highlighting that these had been constructed utilizing the identical expertise as the corporate’s flagship Gemini 2.0 fashions. The Gemma sequence is understood for being open-source and providing on-device efficiency. Google revealed that to date, the Gemma fashions have been downloaded greater than 100 million instances and have been used to create greater than 60,000 variants.

The tech big claimed that Gemma outperforms Meta’s Llama-405B, DeepSeek-V3, OpenAI’s o3-mini AI fashions on the LMArena’s leaderboard. These fashions can be found in 4 sizes — 1B, 4B, 12B, and 27B parameters. They could be run on a single GPU or TPU, the corporate claimed.

Gemma 3 sequence presents superior textual content and visible reasoning capabilities, and may analyse photographs, textual content, and quick movies. The AI fashions provide a context window of 1,28,000 tokens. The fashions additionally provide perform calling assist, which is able to enable builders to construct agentic capabilities within the apps and software program they create.

Google mentioned that the AI fashions had been developed utilizing cautious threat evaluation. The firm claimed that it enforced inside security insurance policies through fine-tuning and benchmark evaluations. The fashions had been additionally examined utilizing extra succesful fashions. The mannequin displayed a low threat stage, the corporate claimed.

In addition to Gemma 3 sequence, the corporate additionally launched ProtectGemma 2, a 4B parameter picture security checker, that ensures that the AI fashions don’t generate content material with harmful, sexually express, or violent content material. Developers have additionally been given the choice to additional customise ProtectGemma to reinforce the security parameters. Gemma 3 household of AI fashions could be downloaded through Google’s Hugging Face listing or on Kaggle.