Google launches Gemma 3, the most powerful AI model you can run on a single GPU
Google has announced the release of Gemma 3, marking the latest advancement in their series of open models. Built on the same research and technology that powers the Gemini 2.0 models, Gemma 3 is designed to operate efficiently across various devices, from phones and laptops to workstations, enabling developers to create AI applications wherever needed.
Gemma 3 is available in a range of sizes — 1B, 4B, 12B, and 27B — offering flexibility to suit different hardware and performance requirements. The models deliver strong performance metrics, surpassing Llama-405B, DeepSeek-V3, and o3-mini in preliminary human preference evaluations.
Supporting over 35 languages out-of-the-box and offering pretrained support for more than 140 languages, Gemma 3 can be utilized for building applications that analyze images, text, and short videos. It also handles complex tasks with an expanded context window and facilitates AI-driven workflows through function calling.
In conjunction with Gemma 3, Google is introducing ShieldGemma 2, a robust 4B image safety checker based on the Gemma 3 framework. Users can explore Gemma 3 in their browser via Google AI Studio, obtain an API key to use with the Google GenAI SDK, download models from platforms like Hugging Face, Ollama, or Kaggle, or deploy them on Vertex AI.
- AI Chatbot
- Paid • Proprietary
Related news
External links
- Gemma 3: Google’s new open model based on Gemini 2.0Google • Official source
- Introducing Gemma 3: The Developer GuideGoogle • Official source
- Google calls Gemma 3 the most powerful AI model you can run on one GPUThe Verge
- Google unveils Gemma 3 as world's best single-accelerator model9to5Google
- Google introduces the Gemma 3 family of accessible lightweight modelsSiliconANGLE
Comments
Google having close ties to the US government this will be used against the people.
Good, but if anything like Gemini, its safeguards will severely limit its creativity and use.
I want a AI with no safeguards or Guardrails.
You can still trick the big models into having no safeguards if you use the API. It's extremely difficult to do it in the chatbots.
Grok 3 has much fewer guardrails. Gab also has even less, but uses weaker models. Then there are some open-source ones on Huggingface that are uncensored if you have a powerful gpu (or rent online).