Google DeepMind has unveiled Gemma 4, its most capable open-source AI model family to date, and the announcement comes with a game-changing surprise: the company has finally switched to the Apache 2.0 license, addressing years of community criticism about restrictive licensing.
A Licensing Revolution
When Google launched Gemma 3, the company faced significant backlash from the open-source community. The custom license restricted commercial use, prohibited competitive deployment, and created legal gray areas that made many developers hesitant to adopt the technology. Gemma 4 changes all of that.
“We listened closely to what innovators need next to push the boundaries of AI, and Gemma 4 is our answer: breakthrough capabilities made widely accessible under an Apache 2.0 license,” wrote Clement Farabet, VP of Research at Google DeepMind, and Olivier Lacombe, Group Product Manager, in the official announcement.
The Apache 2.0 license is one of the most permissive in open-source software. It allows developers to freely use, modify, distribute, and commercialize the technology without the restrictions that plagued previous Gemma releases. This move aligns Gemma 4 with other major open-source projects and removes barriers to enterprise adoption.
Performance That Rivals Closed-Source Giants
Gemma 4 isn’t just about licensing ??it’s a technical powerhouse. The model family ranks #3 globally on the Arena AI text leaderboard, outcompeting models twenty times its size. The 31B Dense variant currently holds the top position among open models, while the 26B Mixture of Experts (MoE) model secured the #6 spot.
The secret sauce? Intelligence-per-parameter. Gemma 4 delivers unprecedented capability density, meaning developers can achieve frontier-level performance with significantly less hardware overhead. For organizations running AI on consumer GPUs or edge devices, this efficiency gain is transformative.
The models come in four versatile sizes:
- E2B (Effective 2B): Optimized for mobile and IoT devices with near-zero latency
- E4B (Effective 4B): Mobile-first AI with multimodal capabilities
- 26B MoE: Mixture of Experts architecture for fast, efficient inference
- 31B Dense: Maximum quality for researchers and developers
Technical Capabilities
Built on the same research and technology as Gemini 3, Gemma 4 brings sophisticated capabilities to the open-source world:
Advanced Reasoning: Multi-step planning and deep logic capabilities make Gemma 4 suitable for complex problem-solving tasks. Math benchmarks and instruction-following tests show significant improvements over previous generations.
Agentic Workflows: Native support for function-calling, structured JSON output, and system instructions enables developers to build autonomous agents that can interact with tools and APIs.
Extended Context: Edge models feature a 128K context window, while larger variants offer up to 256K ??enough to process entire code repositories or lengthy documents in a single prompt.
Multimodal Mastery: All models process video and images natively, supporting variable resolutions and excelling at visual tasks including OCR and chart understanding.
Multilingual Muscle: With native training on over 140 languages, Gemma 4 helps developers build inclusive applications for global audiences.
The Gemma Ecosystem
Since launch, developers have downloaded Gemma over 400 million times, creating what Google calls the “Gemmaverse” ??more than 100,000 custom variants fine-tuned for specific use cases.
Notable examples include INSAIT’s Bulgarian-first language model (BgGPT) and Yale University’s Cell2Sentence-Scale project, which discovered new pathways for cancer therapy using the models.
The ecosystem support is comprehensive: day-one integrations with Hugging Face (Transformers, TRL, Transformers.js, Candle), LiteRT-LM, vLLM, llama.cpp, MLX, Ollama, NVIDIA NIM and NeMo, LM Studio, Unsloth, SGLang, Cactus, Baseten, Docker, MaxText, Tunix, Keras, and Google Cloud services.
Mobile-First AI
For edge deployment, Google partnered closely with the Pixel team and hardware leaders Qualcomm Technologies and MediaTek. The E2B and E4B models run completely offline on phones, Raspberry Pi, NVIDIA Jetson Orin Nano, and other edge devices.
Android developers can prototype agentic flows today using the AICore Developer Preview, with forward-compatibility built-in for Gemini Nano 4.
What This Means for the AI Industry
Google’s licensing shift represents a significant thawing in the relationship between tech giants and the open-source community. By removing restrictive terms, Google positions Gemma 4 as a truly open alternative to proprietary models ??a move that could accelerate AI adoption across industries that previously hesitated due to licensing concerns.
The timing is strategic. With OpenAI raising billions and Meta continuing to push open-source Llama models, Google needed to demonstrate commitment to the developer community. Gemma 4’s Apache 2.0 license is a clear signal that the company intends to compete for open-source mindshare, not just enterprise contracts.
Developers can start experimenting immediately through Google AI Studio (31B and 26B MoE) or AI Edge Gallery (E4B and E2B). Model weights are available on Hugging Face, Kaggle, and Ollama.
The message is clear: the era of restrictive “open” AI models is ending. With Gemma 4, Google is betting that true openness will win in the long run.