Gemma 4: The Next Generation of Open Models

Google has officially unveiled Gemma 4, their most intelligent and capable family of open models to date. Built on the same world-class research as Gemini 3, Gemma 4 is specifically engineered for advanced reasoning and agentic workflows, delivering a breakthrough in "intelligence-per-parameter."

This new release aims to provide developers with frontier-level capabilities without the heavy hardware overhead typically associated with large-scale AI.

1. Four Versatile Model Sizes

Gemma 4 is being released in four distinct configurations to suit various hardware environments and use cases:

• Effective 2B (E2B) & 4B (E4B): Optimized for high-performance edge computing and mobile devices.
• 26B Mixture of Experts (MoE): Focuses on low latency, activating only 3.8B parameters during inference.
• 31B Dense: Maximizes raw quality for deep reasoning and fine-tuning.

2. Advanced Reasoning & Agency

Matching the shift toward autonomous AI, Gemma 4 introduces native support for agentic workflows. This includes:

• Multi-step Planning: Significant improvements in complex logic and mathematical benchmarks.
• Tool-Use: Built-in support for function-calling and structured JSON output.
• Long Context: A massive 128K window for edge models and up to 256K for larger models, allowing for entire repositories to be processed in a single prompt.

3. Multimodal from the Ground Up

Unlike many open models that focus solely on text, Gemma 4 is natively multimodal.

• Vision: All models process video and images with variable resolution, excelling at OCR and chart understanding.
• Audio: The E2B and E4B models include native audio input for real-time speech recognition and understanding.
• Global Reach: Trained natively on over 140 languages for a truly global developer audience.

4. Open and Accessible

In a win for the developer community, Gemma 4 is released under the commercially permissive Apache 2.0 license. This provides complete digital sovereignty, giving developers total control over their data and infrastructure.

The models offer day-one support for industry-standard tools like Hugging Face, vLLM, Ollama, and NVIDIA NIM, and are optimized to run on everything from Raspberry Pi and Android phones to H100 GPUs.

"Gemma 4 represents a new standard for open AI. It isn't just about parameter count; it’s about providing the most efficient reasoning engine possible for every device."

Google's continued commitment to open models ensures that the latest breakthroughs in AI research remain accessible to everyone, from individual hobbyists to enterprise researchers.