Google Unveils Gemma 4: A Leap Forward in Open AI Models
Google has introduced Gemma 4, the latest addition to its family of open AI models, built upon the advanced research and technology that powered Gemini 3. This new release is designed to cater to a wide range of devices, from smartphones to high-performance developer workstations.
Diverse Model Sizes for Varied Applications
Gemma 4 is available in four distinct configurations:
– 31B Dense Model: Ranked as the third leading open model globally on the Arena AI text leaderboard.
– 26B Mixture of Experts (MoE): Holds the sixth position on the same leaderboard.
– Effective 4B (E4B): Optimized for devices requiring efficient performance.
– Effective 2B (E2B): Tailored for ultra-lightweight applications.
In collaboration with the Pixel team, Qualcomm, and MediaTek, Google has fine-tuned the E2B and E4B models to operate seamlessly on devices like smartphones, Raspberry Pi, and Jetson Nano, achieving near-zero latency.
Enhanced Capabilities and Performance
Gemma 4 is engineered to handle tasks ranging from simple conversational interactions to complex logical reasoning and autonomous workflows. Notably, it outperforms models that are 20 times its size. The edge models feature a 128K context window, while the larger variants extend up to 256K, facilitating the processing of extensive documents and datasets in a single prompt.
The models are inherently capable of processing video and images, enabling functionalities such as Optical Character Recognition (OCR) and chart analysis. Additionally, the E2B and E4B models support native audio input for speech recognition and comprehension. Gemma 4 has been trained across more than 140 languages, ensuring broad applicability.
Key Features of Gemma 4
– Advanced Reasoning: Demonstrates significant improvements in multi-step planning and deep logical reasoning, excelling in mathematical computations and instruction-following benchmarks.
– Autonomous Workflows: Offers native support for function-calling, structured JSON output, and system instructions, enabling the development of autonomous agents capable of interacting with various tools and APIs to execute workflows reliably.
– Code Generation: Facilitates high-quality offline code generation, transforming workstations into powerful, local-first AI coding assistants.
Open-Source Accessibility
Gemma 4 is released under the Apache 2.0 license, a commercial and business-friendly open-source license. This licensing grants developers complete control over their data, infrastructure, and models, allowing for flexible development and secure deployment across diverse environments, whether on-premises or in the cloud.
Accessing Gemma 4
Developers can access Gemma 4 through Google AI Studio for the 31B and 26B MoE models, and via the Google AI Edge Gallery for the E4B and E2B models. Model weights are available for download from platforms such as Hugging Face, Kaggle, and Ollama.
With the launch of Gemma 4, Google continues to advance the field of open AI models, providing developers with versatile tools to create innovative applications across a spectrum of devices and use cases.
