Google has launched Gemma 4, a new family of advanced AI models designed for reasoning and agentic workflows.
Google has launched Gemma 4, a new series of open AI models aimed at enhancing advanced reasoning and agentic workflows.
"Gemma 4 — our most intelligent open models to date. Built from the same research as Gemini 3, they bring breakthrough intelligence directly to your hardware for advanced reasoning and agentic workflows," the company announced on Twitter.
The Gemma series has seen significant traction since its initial release, with over 400 million downloads and more than 100,000 model variants created in the Gemmaverse ecosystem. This latest iteration leverages the cutting-edge research that powered the Gemini 3 chatbot.
Model Variants
The Gemma 4 family features four distinct versions: Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts (MoE), and 31B Dense.
The E2B and E4B models are designed for seamless integration and low latency, featuring 2.3 billion and 4.5 billion active parameters, respectively. These models can operate on smartphones and standard laptops.
In contrast, the 26B MoE and flagship 31B models, which have 26 billion and 31 billion parameters, necessitate an Nvidia H100 GPU with 80 GB of memory, targeting researchers and developers. Benchmark results have placed the 31B model third, with the 26B model sixth in the global ranking of open text models by Arena AI, outperforming competitors with significantly larger parameters.
Notable Features
One of the standout features of Gemma 4 is its advanced reasoning ability, enabling the construction of complex logic and the execution of multi-step tasks. The models have shown improvements in mathematical tasks and adhere closely to user instructions.
Additional features include:
- Agentic Workflows — Support for function calls, structured JSON output, and system instructions, facilitating the creation of autonomous assistants that can interact with APIs;
- Code Generation — High-quality code writing capabilities in offline mode, transforming workstations into local AI assistants;
- Vision and Audio Processing — All models can handle video and image processing, text recognition, and chart analysis, with E2B and E4B also supporting speech recognition;
- Extended Context Window — The compact versions support up to 128,000 tokens, while the larger models can manage up to 256,000 tokens, suitable for processing extensive documents in a single query;
- Multilingual Capability — The models support over 140 languages.
Gemma 4 is available in Google AI Studio and Google AI Edge Gallery and can be integrated with various third-party tools and frameworks, including Hugging Face and NVIDIA NIM. Customization options are available via Google Colab, Vertex AI, or local GPUs, with deployment options on Google Cloud services.
This announcement follows Google’s introduction of the Veo 3.1 Lite AI model for video generation earlier this month.
