Key Takeaways
- Google DeepMind released Gemma 4 on April 2, 2026, a family of four free, open-source AI models built from its Gemini 3 research and available under the permissive Apache 2.0 license for the first time.
- The four models (Effective 2B, Effective 4B, 26B MoE, and 31B Dense) are designed to run locally on smartphones, laptops, Raspberry Pi devices, and single GPUs with no cloud subscription or per-use fees required.
- All four models support image and video processing, with the two smallest also handling native audio input, and context windows reach up to 256,000 tokens for the larger models.
Google Launches Gemma 4: What You Need to Know
Google DeepMind launched Gemma 4 on Wednesday, April 2, 2026, releasing what it describes as the most capable open-weight AI model family it has ever built. The models are available immediately under an Apache 2.0 license, which is the most permissive open-source license Google has used for Gemma to date. In plain terms, that means developers and businesses can use, modify, and build commercial products with these models without restrictive terms.
For everyday users, the big takeaway is simple: Gemma 4 lets powerful AI run directly on your own devices, including phones, laptops, and even a $60 Raspberry Pi, without needing an internet connection or paying monthly cloud fees.
The Four Gemma 4 Model Sizes
The Gemma 4 models come in four sizes: Effective 2B, Effective 4B, a 26B Mixture of Experts model, and a 31B Dense model.
The smaller Effective models are designed for edge use cases on lightweight hardware such as Android smartphones or Raspberry Pi computers. Google's LiteRT-LM runtime allows the Effective 2B model to run using under 1.5 gigabytes of memory on supported devices. To put that in perspective, most modern smartphones have 6 to 12 GB of RAM, so the smallest Gemma 4 model uses just a fraction of what your phone already has.
The 26B MoE model only activates 3.8 billion parameters during inference tasks, allowing it to perform at high speed without sacrificing the deep knowledge base of larger models. The 31B Dense model is the most powerful of the four and is aimed at developers running workstations or dedicated GPUs. The unquantized weights fit on a single 80GB NVIDIA H100 GPU, and quantized versions run on consumer-grade GPUs for local coding assistants and development workflows.
Performance Benchmarks
The 31B Dense variant currently ranks third among open models on the industry-standard Arena AI Text leaderboard. The 31B Dense model achieved an estimated LMArena text score of 1452, while the 26B MoE reached 1441 with just 4 billion active parameters. For context, these benchmark scores place Gemma 4 in direct competition with much larger models from other companies.
What Can Gemma 4 Actually Do?
Multimodal Capabilities
All four models have the ability to process images and videos, with the smaller E2B and E4B variants going further with support for native audio inputs, enabling real-time speech understanding directly on device. This means the smallest models can listen to and understand spoken language without sending your audio to any server.
AI Agents and Function Calling
The Gemma 4 models have native support for function calling and structured JSON outputs, meaning developers can use them to power autonomous agents that interact with third-party tools and execute multi-step plans. In non-technical terms, these models can be programmed to take actions on your behalf, such as searching databases, filling out forms, or managing multi-step tasks, all running locally on your hardware.
Context Window
Google has increased the context window up to 128K tokens for the smallest models and 256K for the larger two. A "context window" is how much text the AI can process at once. At 256K tokens, you could feed the model an entire novel or a full software codebase in a single prompt.
Language Support
Gemma 4 supports over 140 languages, making it one of the most linguistically diverse open-source AI model families available.
Where to Get Gemma 4
The model weights are available for download from Hugging Face, Kaggle, and Ollama. Gemma 4 runs on Android and iOS with CPU and GPU support, and on Windows, Linux, and macOS. Browser-based execution through WebGPU is also supported.
Google is also launching a new Python package and command-line tool that lets developers experiment with Gemma 4 without writing any code.
Apache 2.0 License: Why It Matters
Previous Gemma releases used Google's own custom license, which included restrictions on certain commercial uses. The Apache 2.0 license removes many of the commercial restrictions that have made other open models less attractive for enterprise use. This is a significant shift. Apache 2.0 is widely trusted in the software industry and gives developers and companies full freedom to use, modify, and distribute the models in commercial products without special permission.
How Gemma 4 Compares to the Competition
The open-source AI model space is crowded. Meta's Llama 4 and Mistral have been the go-to options for developers in the open-source AI space, and Google has spent years trying to close that gap. With Gemma 4's strong benchmark performance, permissive licensing, and small-device support, Google is making its most direct play for developer adoption.
It is worth noting that Gemma 4 is not the same as Google's Gemini models. Gemini is Google's proprietary, cloud-based AI (used in products like Google Search and the Gemini chatbot). Gemma is the open-source cousin, built from the same underlying research but designed to run independently on your own hardware.
What This Means for Regular Users
If you are not a developer, Gemma 4 still matters. As device manufacturers and app developers adopt these models, you will likely see smarter AI features built into phones, smart home devices, and desktop software that work without an internet connection. Voice assistants, offline translation tools, and on-device photo analysis could all improve as a result of models like Gemma 4 becoming freely available.
For home networking enthusiasts and tech-forward consumers, the ability to run capable AI locally also means more processing stays on your own network rather than being routed through external servers, which has privacy and latency benefits.
FAQ
What is Google Gemma 4?
Gemma 4 is a family of four free, open-source AI models released by Google DeepMind on April 2, 2026. Built from the same research behind Google's Gemini 3, the models are designed to run locally on phones, laptops, Raspberry Pi devices, and GPUs without needing cloud services or subscriptions.
Is Gemma 4 free to use?
Yes. Gemma 4 is released under the Apache 2.0 open-source license, which means it is free to download, use, modify, and build commercial products with. There are no per-use fees or subscription costs.
What is the difference between Gemma 4 and Google Gemini?
Gemini is Google's proprietary, cloud-based AI used in products like Google Search and the Gemini chatbot app. Gemma is the open-source counterpart built from similar research but designed to run independently on personal devices without requiring Google's cloud infrastructure.
Can Gemma 4 run on a smartphone?
Yes. The smallest Gemma 4 model (Effective 2B) uses under 1.5 GB of memory and is designed to run on Android smartphones, iPhones, and even Raspberry Pi computers. It supports CPU and GPU processing on mobile devices.
What sizes does Gemma 4 come in?
Gemma 4 comes in four sizes: Effective 2B and Effective 4B for mobile and edge devices, a 26B Mixture of Experts model for fast performance with low resource use, and a 31B Dense model for maximum quality on workstations and GPUs.
Does Gemma 4 require an internet connection?
No. Once downloaded, Gemma 4 models run entirely on your local device. No internet connection, cloud subscription, or external server is required, which also means your data stays private on your own hardware.
How does Gemma 4 compare to Meta Llama 4 and other open-source AI models?
Gemma 4's 31B Dense model ranks third among open models on the Arena AI Text leaderboard. It competes directly with Meta's Llama 4 and Mistral models, with advantages in small-device support, multimodal capabilities (image, video, and audio), and the permissive Apache 2.0 license.
