Overview
Google Gemma is a family of lightweight, open-source AI models built by Google DeepMind, designed to bring frontier-level intelligence to devices ranging from smartphones to servers. Released under the Apache 2.0 license with full commercial freedom, Gemma gives developers and researchers direct access to the same research and technology used to build Google's Gemini models, with open weights released under Apache 2.0 and no software licensing fee for self-hosting.
Gemma targets AI developers, researchers, and enterprises who need high-performance language models they can run locally, fine-tune for specific use cases, or deploy at the edge without relying on cloud APIs. With the latest Gemma 4 release (April 2026), the family spans four model sizes from 2B to 31B parameters, with text and image input across all variants, plus video understanding via frames; native audio input is available only on the E2B and E4B edge variants.
The model family matters because it democratizes access to advanced AI agent capabilities. Gemma 4's 31B Dense model ranks as the #3 open model on the Arena AI text leaderboard, while its edge variants run entirely offline on phones and single-board computers with near-zero latency — making it practical for privacy-sensitive, low-connectivity, and resource-constrained environments.
Key Features
Multimodal Understanding — Processes text, images, video (up to 60 seconds), and audio input natively across all model sizes, enabling applications that reason across different media types without separate pipelines.
Agentic Workflow Support — Purpose-built for agentic workflows, with native function calling, structured JSON output, and system instructions for building multi-step task agents.
Edge-Optimized Inference — The E2B and E4B variants activate only 2–4 billion parameters during inference, preserving RAM and battery life while running completely offline on phones, Raspberry Pi, and NVIDIA Jetson devices.
Extended Context Windows — Supports up to 128K tokens on edge models and 256K tokens on larger variants, allowing processing of long documents, codebases, and extended conversations without truncation.
Multilingual Coverage — Gemma 4 is pretrained on 140+ languages and offers out-of-the-box support for 35+ languages. In Google's official benchmarks, the 31B model scores 88.4% on MMMLU, making it suitable for global applications and multilingual AI text generator use cases.
Fine-Tuning Flexibility — Compatible with popular training frameworks including JAX, Keras, and Hugging Face Transformers, so teams can adapt models to domain-specific tasks — from AI data science pipelines to custom chatbots — with their existing toolchains.
How to Get Started
Getting started with Gemma depends on your deployment target and preferred workflow.
Local Desktop/Laptop (Quickest)
Download and run Gemma through Ollama with a single command:
ollama run gemma4:26b
This handles model download, quantization, and serving automatically. For the edge-optimized variant, use gemma4:e4b.
Cloud & Production Deployment
- Google Cloud Vertex AI — Deploy Gemma 4 with managed infrastructure, auto-scaling, and monitoring through the Vertex AI Model Garden
- Google Kubernetes Engine — Run containerized Gemma workloads with GPU acceleration using the official Docker images
- vLLM — High-throughput serving for production API endpoints with OpenAI-compatible interfaces
Research & Fine-Tuning
- Hugging Face — Access all model variants with Transformers integration for fine-tuning and evaluation
- Kaggle — Run experiments directly in notebooks with free GPU access
- LM Studio — Desktop application for local experimentation with a visual interface
All models are available for immediate download from Hugging Face, Kaggle, and the Google AI Developer portal.
Pricing & Plans
Gemma is completely free and open source under the Apache 2.0 license.
Open Source (Free)
- All model weights available for unrestricted download
- No monthly active user caps
- No acceptable-use restrictions beyond Apache 2.0 terms
- Full commercial use permitted — no royalties, no registration
- Self-host on your own infrastructure at zero software cost
Third-Party API Access
- Google Cloud Vertex AI, OpenRouter, and other providers offer hosted Gemma endpoints
- Hosted API pricing varies by provider and changes over time; check each provider's current rate card
- Cloud provider costs depend on instance type and usage volume
- Free-tier GPU access available through Kaggle and Google Colab for experimentation
Hardware Requirements
- E2B/E4B: Runs on smartphones, Raspberry Pi, and devices with 4–8 GB RAM
- 26B A4B: Positioned by Google for desktop computers and small servers
- 31B Dense: Positioned for large servers or server clusters; consult Google's deployment guide for hardware specifics
Community & Ecosystem
Gemma benefits from one of the broadest open-model ecosystems available. As an Apache 2.0 project, it has attracted rapid adoption across platforms and toolchains.
Platform Support
- Hugging Face: Full integration with Transformers, including model cards, discussion forums, and community fine-tunes
- Ollama: One-command local deployment with automatic quantization
- LM Studio: Desktop GUI for non-technical users to experiment locally
- NVIDIA: Day-one optimization for RTX GPUs through the AI Garage program, with TensorRT-LLM acceleration
Framework Compatibility
- JAX, Keras, PyTorch, and Hugging Face Transformers for training and fine-tuning
- vLLM, llama.cpp, and TGI for high-performance serving
- Google AI Edge SDK for mobile and IoT deployment
- Google AI Edge SDK and MediaPipe LLM Inference API for native Android integration
Developer Resources
- Official documentation at ai.google.dev/gemma
- Model cards with detailed benchmark data and safety evaluations
- Cookbook examples for common use cases (RAG, function calling, agentic workflows)
Best For
- AI developers building privacy-first applications that must run entirely on-device without cloud dependencies
- Researchers who need an open, fine-tunable foundation model with strong multilingual and multimodal capabilities
- Startups looking for a commercially permissive AI model to integrate into products — including AI app builders — without licensing overhead
- Mobile and IoT developers deploying intelligent features on resource-constrained edge hardware
- Enterprise teams running air-gapped or compliance-sensitive environments where data cannot leave the network
FAQ
Is Google Gemma really free to use commercially?
Yes. Gemma is released under the Apache 2.0 license, which allows unrestricted commercial use, modification, and redistribution. There are no monthly active user caps, no registration requirements, and no royalty payments. You can integrate Gemma into commercial products, fine-tune it on proprietary data, and distribute modified versions freely.
What is the difference between Gemma and Gemini?
Gemma is Google's open-source model family designed for local deployment and fine-tuning. Gemini is Google's proprietary, cloud-hosted model powering products like Google AI Studio and Gemini in Chrome, including AI chatbots and conversational interfaces. Gemma is built using similar research and technology as Gemini but is smaller, open-weight, and optimized for running on developer hardware rather than Google's data centers.
Can Gemma 4 run on my phone?
Yes. The E2B and E4B edge variants are specifically designed for mobile deployment. They activate only 2–4 billion parameters during inference, using minimal RAM and battery. Google has not documented Gemma 4 as running inside Android's AICore system service. Official mobile deployment paths for Gemma include Google AI Edge Gallery and the MediaPipe LLM Inference API, while AICore is documented for Gemini Nano.
How does Gemma 4 compare to Llama and Mistral?
Gemma 4's 31B model ranks #3 on the Arena AI text leaderboard among open models, competing directly with Meta's Llama and Mistral's offerings. Its key differentiators are native multimodal support across all sizes, Apache 2.0 licensing (vs. Llama's custom license), and edge-optimized variants that run on devices as small as a Raspberry Pi.
What hardware do I need to run Gemma 4 locally?
For the E2B/E4B edge models: a smartphone or device with 4–8 GB RAM. For the 26B MoE model: a machine with 16+ GB RAM using quantized weights. For the full 31B Dense model: a GPU with 24+ GB VRAM (like an RTX 4090) or a machine with 20+ GB RAM for CPU inference at reduced speed.
Does Gemma support function calling and tool use?
Yes. Gemma 4 is purpose-built for agentic workflows with native support for function calling, tool use, and multi-step planning. It scores 86.4% on the t2-bench agentic tools benchmark, making it suitable for building autonomous agents that interact with external APIs and applications.
What languages does Gemma 4 support?
Gemma 4 is pretrained on over 140 languages and provides out-of-the-box support for 35+ languages. In Google's official benchmark table, the 31B model scores 88.4% on MMMLU. This breadth makes it suitable for applications targeting diverse international audiences.
Can I fine-tune Gemma on my own data?
Yes. Gemma supports fine-tuning through standard frameworks including Hugging Face Transformers, JAX, Keras, and PyTorch. Google provides official fine-tuning guides and cookbook examples. You can adapt any Gemma variant to your domain using techniques like LoRA, QLoRA, or full fine-tuning depending on your hardware resources.
