On March 12, 2025, Google announced Gemma 3 , an open source,
commercially available large-scale language model. According to Google, Gemma 3 is the world’s best large-scale language model that can run on a single GPU or TPU.
Gemma 3: Google’s new open model based on Gemini 2.0
https://blog.google/technology/developers/gemma-3/
Gemma 3 model overview | Google AI for Developers
https://ai.google.dev/gemma/docs/core
Gemma3Report.pdf
(PDF file) https://storage.googleapis.com/deepmind-media/gemma/Gemma3Report.pdf
Google has released ‘ Gemma ‘ as an open source large-scale language model that utilizes Gemini’s research resources, and in May 2024 released ‘Gemma 2,’ a large-scale language model that can deliver performance comparable to Llama 3 70B, which has 70 billion parameters.
Google releases open source visual language model ‘PaliGemma’ and announces large-scale language model ‘Gemma 2’ with performance equivalent to Llama 3 – GIGAZINE
Now, Google has announced Gemma 3, an enhanced version of Gemma 2. Built using the same research and technology as Google’s large-scale language model Gemini 2.0 , Gemma 3 is said to be ‘the most advanced, portable, and responsibly developed open source model to date.’
Gemma 3 has four models with 1 billion, 4 billion, 12 billion, and 27 billion parameters, allowing users to choose the best model for their specific hardware and performance needs. In addition, Google says, ‘Gemma 3 is designed to run fast on any device, including smartphones, laptops, and workstations, so developers can create AI applications wherever they want.’
One of the features of Gemma 3 is that it can run on a single GPU or TPU. In the ‘ Chatbot Arena ,’ which is used to measure the relative capabilities of large-scale language models, it has been reported that the Gemma 3 27B, which uses a single NVIDIA H100, can outperform DeepSeek v3 and Llama 3 405B, which use multiple H100s.
Gemma 3 is also pre-trained in over 140 languages, allowing users to build applications that are tailored to their language needs, and has a context window of 128,000 tokens, allowing it to handle complex tasks.
‘In developing Gemma 3, we made some fine-tuning changes to ensure it is aligned with our broader data governance and safety policies. As we develop more powerful models, it is important that the industry collaborates to develop an approach to safety that is proportionate to the risks. We will continue to learn and improve safety measures over time in open source models like Gemma 3,’ Google said.
Gemma 3 is available on
Kaggle and Hugging Face.
Gemma 3 Release – a google Collection
https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d
Gemma 3 can also be run on
Ollama , a library that allows you to run large language models locally.
gemma3
https://ollama.com/library/gemma3