Gemma 4: Google's Open-Source AI Models
On April 1, 2026, Google unveiled Gemma 4. Four models. Apache 2.0 license. Architecture derived from Gemini 3.
It's not an incremental update. It's a change in strategy: frontier-level models distributed with no commercial restrictions.
Four Models for Four Scenarios
The Gemma 4 family is designed for radically different environments. Each variant has a precise target.
Effective 2B (E2B)
Built for edge devices. Smartphones, IoT systems, embedded hardware like Raspberry Pi. It's the lightest model in the family.
Effective 4B (E4B)
Advanced mobile computing. Developed in collaboration with Qualcomm and MediaTek. Optimized for NPUs and next-generation mobile chips.
26B Mixture of Experts (MoE)
MoE architecture: activates only 4 billion parameters per inference, regardless of the model's total size. The computational cost is low. The performance isn't.
Currently sixth in LMArena's text leaderboard.
31B Dense
The family's flagship model. Designed for workstations and local servers. Third place globally on LMArena with a score of about 1,452 points.
All Gemma 4 models support multimodal input: text, images, audio, and video. They're optimized for agentic workflows, code generation, and advanced system prompting.
Performance: The Numbers That Matter
The 31B beats Gemini 2.5 Pro and Qwen 3.5-397B on public benchmarks. It comes close to the performance of Claude 4.5 Sonnet.
The most interesting figure isn't the absolute score. It's the capability density per parameter. According to Google, computational efficiency is comparable to models up to 30 times larger.
On the hardware front: both the 26B and the 31B run on a single 80GB Nvidia H100 with unquantized weights. For anyone working with on-premise infrastructure, that's the number that changes the math.
The Apache 2.0 License: Why It Matters
The real break isn't technical. It's legal.
Gemma 4 adopts the Apache License 2.0. That means anyone can:
- use the models
- modify them
- distribute them
- integrate them into commercial products
With no substantial restrictions. No royalties. No usage constraints.
For many observers, this choice will have a bigger impact than the benchmarks. A model with a permissive license enters products, pipelines, and companies differently than a closed model or one with limited commercial-use terms.
The Context: Competition Among Open Models
The launch arrives at a moment of acceleration in the open-model space.
In recent months, Chinese competitors have gained ground. Alibaba with its Qwen series. DeepSeek with its own models. Both are betting on efficiency and free distribution.
Google is responding with Gemma 4 and significant adoption numbers. The Gemma community has reached:
- over 400 million downloads
- more than 100,000 variants created by developers
Among the most notable projects: language models for underrepresented languages (Bulgarian, among others) and Cell2Sentence-Scale, developed at Yale University for cancer research.
These numbers point to one thing: Gemma isn't an academic experiment. It's production infrastructure.
Where to Download Gemma 4
The models are available across multiple platforms. The choice depends on your working environment.
- Google AI Studio — 31B, 26B MoE. Cloud inference, no setup.
- Google AI Edge Gallery — E2B, E4B. Android/iOS app for edge devices.
- Hugging Face — All models. Standard for research and fine-tuning.
- Ollama — All models. Local execution via CLI.
- Kaggle — All models. Notebooks and research.
- Nvidia NGC — 31B, 26B MoE. Optimized for Nvidia GPUs.
- Android Studio — E2B, E4B. Local AI coding via the AICore Developer Preview.
The models are optimized for Nvidia GPUs, AMD GPUs, and Google Cloud TPUs.
FAQ on Gemma 4
What is Gemma 4 and who developed it?
Gemma 4 is a family of open-source language models developed by Google and released on April 1, 2026. It's built on the same architecture as Gemini 3 and distributed under an Apache 2.0 license.
What are the differences between the Gemma 4 models?
The family includes four variants: E2B for edge devices, E4B for mobile computing, 26B MoE for efficient inference, and 31B Dense for workstations and servers. Each model has a different hardware and performance profile.
Can I use Gemma 4 for commercial applications?
Yes. The Apache 2.0 license allows commercial use, modification, and distribution with no substantial restrictions.
How does Gemma 4 compare to other open-source models?
The 31B model holds third place on LMArena's text leaderboard with about 1,452 points, beating Gemini 2.5 Pro and Qwen 3.5-397B. The 26B MoE is sixth, despite using only 4B active parameters per inference.
What hardware do you need to run Gemma 4 locally?
The 26B and 31B models run on a single 80GB Nvidia H100 with unquantized weights. The E2B and E4B models are designed for mobile and edge devices, including Raspberry Pi.
Conclusion
Gemma 4 brings competitive-level models to accessible hardware. Under an Apache 2.0 license, with no commercial constraints.
The advantage isn't just in the benchmarks. It's in the freedom of integration. Anyone building AI-powered products, managing on-premise infrastructure, or developing vertical applications now has a solid, free option.