What Are the Improvements of Gemma 3 over Gemma 2? 196

Just now, Google released Gemma 3, available in four parameter sizes — 1B, 4B, 12B, and 27B — and runnable on a single GPU/TPU! With just 27B parameters, Gemma 3 has defeated the DeepSeek 671B model, making it the best open-source model second only to DeepSeek R1. Compared to Gemma 2, Gemma 3 has been improved in the following ways:

Multimodal Capabilities

Gemma 3 introduces support for mixed input of text, images, and short videos, capable of handling complex multimodal tasks such as image question-answering and video content analysis. In contrast, Gemma 2 was mainly focused on text processing capabilities.

Performance Improvements

Gemma 3 is called "the world's strongest single-accelerator model," significantly outperforming Gemma 2 in single GPU or TPU environments, with inference speeds improved by 47% when processing short video content.

Enhanced Language Support

Gemma 3 is pretrained with support for over 140 languages, with direct support for over 35 languages. By comparison, Gemma 2 had a narrower language support range.

Expanded Context Window

Gemma 3's context window has been expanded to 128k tokens, enabling it to handle much larger amounts of information, such as 30 high-resolution images or 1 hour of video content. This is a significant improvement over Gemma 2's context window.

Visual Processing Capabilities

Gemma 3 introduces dynamic image slicing technology combined with frame sampling and optical flow analysis, supporting high-resolution and non-square images, and capable of completing key frame extraction from 1 hour of video within 20 seconds. In contrast, Gemma 2 had weaker visual processing capabilities.

Enhanced Safety

Gemma 3 is equipped with the ShieldGemma 2 image safety classifier, capable of detecting and flagging dangerous content, pornographic content, and violent content, further improving the model's safety.

Hardware Optimization

Gemma 3 has been deeply optimized for Nvidia GPUs and Google Cloud TPUs, ensuring efficient operation across different hardware platforms. In addition, it supports multiple deployment options, including Vertex ai, Cloud Run, and local environments.

Training and Fine-tuning Improvements

Gemma 3 employs techniques such as knowledge distillation, reinforcement learning (including human feedback and machine feedback), and model merging to improve the model's capabilities in mathematics, coding, and instruction following. At the same time, it provides more flexible fine-tuning tools, making it convenient for developers to customize according to their needs.

Community and Ecosystem Expansion

The release of Gemma 3 further enriches the "Gemmaverse" ecosystem, allowing developers to build more innovative applications based on its powerful features.

Overall, Gemma 3 has shown significant improvements in multimodal processing, performance, language support, visual processing, and safety, making it a powerful AI model better suited for deployment and use across a wide range of devices.