According to Zhitong Finance APP, Google (GOOGL.US) released its latest open-source model, Gemma 3, on Wednesday, stating that the model can run on a single GPU or TPU.
Google stated that Gemma 3 is a series of lightweight, state-of-the-art open-source models built on the same research and technology as the Gemini 2.0 model.
Google CEO Sundar Pichai mentioned in a post on X: &#34;Our new open-source model is highly efficient—the largest 27B model can run on just one H100 GPU. Other models would require at least 10 times the computational resources to achieve similar performance.&#34;
According to the company, Gemma 3 will offer four sizes: 1B, 4B, 12B, and 27B, allowing users to choose the best model based on their hardware and performance needs.
Clement Farabet, Vice President of Research at Google DeepMind, and Director Tris Warkentin stated in a blog post on Wednesday: &#34;These models are designed to run directly on edge devices—from smartphones and laptops to workstations, helping developers create AI applications where they are needed.&#34;
<img src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250312/1741786567289690.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="image.png" original-src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250312/1741786567289690.png"/>
Google pointed out that in the preliminary human preference evaluation of the LMArena leaderboard, Gemma 3 outperformed Meta Platforms (META.US) Llama-405B, DeepSeek-V3, and OpenAI's o3-mini.
According to the rankings, only Deepseek's R1 model scored higher than Gemma 3.
Gemma 3 offers out-of-the-box support for over 35 languages and provides pre-training support for more than 140 languages.
Additionally, Google launched ShieldGemma 2, an image safety checker based on the Gemma 3 foundation. According to the company, ShieldGemma 2 provides a ready-made solution for image safety, outputting three types of safety labels—dangerous content, pornographic content, and violent content.
Google also introduced the Gemma 3 academic program, allowing academic researchers to apply for Google Cloud credits (each worth $10,000) to advance their research based on Gemma 3. The application form will be open for four weeks

GOOGL

GOOG

GGLS

GGLL

On Wednesday, Google released its latest open-source model Gemma 3, stating that the model can run on a single GPU or TPU

- Google has released its latest open-source model, Gemma 3, which can run on a single GPU or TPU.  
- The model is designed for efficiency, with the largest 27B version requiring only one H100 GPU, significantly less than other models for similar performance.  
- Gemma 3 supports over 35 languages and includes a new image safety checker, ShieldGemma 2, while also offering academic researchers Google Cloud credits to support their research.

Google's latest AI model Gemma 3 is online! It can run 27B on a single card, surpassing o3-mini