New Gemma 4 models demonstrate expanded multimodal understanding and enhanced efficiency, pushing boundaries for AI development across personal devices and cloud platforms.
Google DeepMind has rolled out its latest suite of Gemma 4 models, marking a significant progression in open-weight large language model technology. The released models, including Gemma 4 31B IT Thinking, Gemma 4 26B A4B IT Thinking, Gemma 4 E4B IT Thinking, and Gemma 4 E2B IT Thinking, showcase enhanced capabilities in processing not only text but also images, video, and audio.
Expanded Multimodal Horizons
The Gemma 4 family distinguishes itself through its expanded multimodal functionalities. Across all model sizes, the systems exhibit support for variable aspect ratio and resolution images. Specifically, the E2B and E4B variants offer native audio and video processing, broadening their application scope considerably. This advancement positions Gemma 4 for more complex tasks, including rich audio-visual understanding and agentic tool use, as demonstrated by its performance on benchmarks like τ2-bench for agentic tool use.
Read More: Huawei focuses on watches and phones due to chip limits
Performance and Safety Gains
"Gemma 4 models significantly outperform Gemma 3 and 3n models in improving safety, while keeping unjustified refusals low."
Evaluations reveal substantial improvements in safety metrics compared to previous Gemma iterations. The models produced minimal policy violations in text-to-text and image-to-text tasks. This focus on safety, coupled with maintained or improved performance on diverse datasets such as MMMLU (Multilingual Q&A) and AIME 2026 Mathematics, underscores a commitment to responsible AI development.
Accessibility and Deployment
The Gemma 4 models are designed for flexible deployment. A JAX library, available on GitHub, facilitates using and fine-tuning the models on personal hardware, including CPUs, GPUs, and TPUs. Furthermore, Gemma 4 is now accessible on Google Cloud, integrated with services like Vertex AI, Google Kubernetes Engine (GKE), and Google Compute Engine (GCE), offering developers robust options for scaling their AI applications. The integration with Google ADK (AI Development Kit) also enables the creation of fully functional AI agents.
Development and Training
The underlying strength of Gemma 4 stems from the quality and diversity of its training data. While specific details on the training dataset remain largely undisclosed, the model card highlights its extensibility for building autonomous agents capable of planning, navigating applications, and completing tasks via native function calling support.
