**Published:** April 28, 2026 | **Category:** AI Open Source Models
## Introduction
Google DeepMind released Gemma 4 on April 2, 2026, under the Apache 2.0 license—the most permissive in the Gemma family history. This release includes four models, each designed for specific use cases from edge devices to high-performance computing.
## The Model Family
### Gemma 4 2B and 4B (Edge Models)
These compact models are optimized for phones and edge devices:
– Native audio input support
– Text, image, and video understanding
– Runs efficiently on mobile hardware
### Gemma 4 26B (Mixture of Experts)
The MoE variant offers a balance of capability and efficiency:
– Only 3.8B parameters activate at inference
– Low-latency performance
– Can run on a single high-end GPU
### Gemma 4 31B (Dense Model)
The flagship model achieving remarkable benchmarks:
– Third place globally among open models on Arena AI
– Elo score of 1452
– 89.2% on AIME 2026
– 80.0% on LiveCodeBench v6
## Technical Specifications
| Model | Parameters | Context Window | Multimodal |
|——-|————|—————|
| 2B Edge | 2B | 32K | Text, Image, Audio |
| 4B Edge | 4B | 32K | Text, Image, Audio |
| 26B MoE | 26B (3.8B active) | 256K | Text, Image, Video |
| 31B Dense | 31B | 256K | Text, Image, Video |
## Day-One Framework Support
Gemma 4 launched with support across all major ML frameworks: Hugging Face Transformers, Ollama, vLLM, llama.cpp, MLX (Apple Silicon), LM Studio, NVIDIA NIM, and Android Studio.
## Getting Started
For Ollama users, Gemma 4 is available now:
“`bash
ollama run gemma4:27b
“`
## Conclusion
Google Gemma 4 represents a significant step forward for open-source AI. The combination of strong performance, permissive licensing, and comprehensive framework support makes these models accessible to developers across all contexts.
**Rating:** 4.7/5