Introduction to Gemma 3
Google has introduced Gemma 3, which it claims is the "world’s best single-accelerator model." This model comes in various sizes, ranging from a 1 billion-parameter model that can run on almost any device to a 27 billion-parameter version that requires significant RAM. The model is also available in 4 billion and 12 billion versions. The smallest Gemma 3 model can occupy less than a gigabyte of memory in lower-precision modes, while the larger versions require 20GB-30GB even at 4-bit precision.
Performance of Gemma 3
Google has provided data that shows significant improvements in Gemma 3’s performance compared to other models. Using the Elo metric, which measures user preference, Gemma 3 27B outperforms Gemma 2, Meta Llama3, OpenAI o3-mini, and other chat models. Although it doesn’t quite match DeepSeek R1 in this subjective test, it runs on a single Nvidia H100 accelerator, whereas most other models require multiple GPUs. Google also claims that Gemma 3 is more capable in math, coding, and following complex instructions, but it doesn’t provide numbers to support this claim.
Chatbot Arena ELO Score
The subjective user preference Elo score shows that people prefer Gemma 3 as a chatbot. This is evident from the chart below, which compares the performance of Gemma 3 with other models.
Availability and Accessibility
The latest Gemma model is available online in Google AI Studio. Users can also fine-tune the model’s training using tools like Google Colab and Vertex AI or use their own GPU. The new Gemma 3 models are open-ish, and users can download the whole model for free from repositories like Kaggle or Hugging Face. However, Google’s license agreement limits what users can do with the models. Despite this, Google won’t be able to track what users are exploring on their own hardware, which is the advantage of having more efficient local models like Gemma 3.
Community and Applications
No matter what users want to do, there’s a Gemma model that will fit on their hardware. For inspiration, Google has a new "Gemmaverse" community to highlight applications built with Gemma models.
Conclusion
Gemma 3 is a powerful and versatile model that offers significant improvements in performance and efficiency. Its availability in various sizes and its open-ish nature make it accessible to a wide range of users. With its potential applications in chat, math, coding, and more, Gemma 3 is an exciting development in the field of AI.
FAQs
- What is Gemma 3?
Gemma 3 is a single-accelerator model developed by Google that comes in various sizes and offers significant improvements in performance and efficiency. - Where can I access Gemma 3?
The latest Gemma model is available online in Google AI Studio, and users can also download it for free from repositories like Kaggle or Hugging Face. - What are the limitations of Gemma 3?
Google’s license agreement limits what users can do with the Gemma 3 models, but users can still explore and use the models on their own hardware without Google’s tracking. - What is the Gemmaverse community?
The Gemmaverse community is a platform where users can find inspiration and showcase applications built with Gemma models.
