Gemma 3: From zero to hero
NotebookLM shared with system prompt and other contexts
Click Crash Courses for grounding sources in NotebookLM
Gemma is a family of lightweight, state-of-the-art open-weights AI models developed by Google DeepMind and other teams across Google. Built using the same research, technology, and infrastructure as Google’s flagship Gemini models, Gemma is designed to offer high-performance AI capabilities that can run locally on consumer hardware like laptops, smartphones, or IoT devices. [1, 2]
Unlike proprietary models accessed strictly via internet APIs, Google releases Gemma under a permissive open license (Apache 2.0), allowing developers and researchers to download the model weights for free, run them completely offline, and customize or fine-tune them for personal or commercial applications. [3, 4, 5]
🌟 Key Features of the Gemma Ecosystem
- On-Device & Local Power: Optimized to run locally without internet connection, ensuring total data privacy and zero latency.
- Multimodal Intelligence: Processes and understands text, images, and audio natively right on your device.
- Configurable Thinking Modes: Supports advanced reasoning capabilities for multi-step math problems or deep programming logic.
- Massive Context Windows: Features large context windows ranging from 128K up to 256K tokens, allowing the model to process massive documents or codebases at once.
- Agentic Workflows: Architected with native function-calling and tool-use capabilities to power autonomous AI agents. [3, 4, 6, 7, 8]
📦 The Gemma Model Family
The lineup includes diverse parameter sizes tailored for specific hardware configurations, ranging from ultra-mobile deployment to enterprise-grade AI servers: [3]
- Gemma Core Models: Available in distinct configurations:
- E2B & E4B (Edge Sizes): Highly efficient “Effective Parameter” versions designed for smartphones (iOS/Android), browsers, and single-board computers like Raspberry Pi.
- 12B (Unified): A mid-sized model utilizing an innovative encoder-free architecture that blends text, image, and native audio directly inside the model backbone.
- 26B A4B (Mixture-of-Experts): A high-throughput version activating only 4 billion parameters per token to balance speed and intelligence.
- 31B (Dense): The flagship powerhouse optimized for local developer workstations, coding assistants, and local enterprise environments. [3, 9, 10, 11, 12]
- Specialized Variants:
- ShieldGemma: Safety classifier models designed to filter and detect policy-violating prompts or completions.
- EmbeddingGemma: Generates high-quality vector embeddings for semantic search and clustering.
- DiffusionGemma: Built to maximize generation and text diffusion speed. [1, 13, 14, 15, 16]
🛠️ How to Access and Run Gemma
Depending on your level of technical expertise, you can run or experiment with Gemma through a few different avenues:
1. Instant Cloud Testing (No Installation) [17]
- Google AI Studio: You can instantly prompt, test, and prototype Gemma models for free directly inside your browser by visiting Google AI Studio.
- Gemini API Integrations: Developers can quickly test or serve the larger Gemma architectures via the Google GenAI SDK. [7, 8, 18, 19, 20]
2. Local Offline Execution (For Workstations & Laptops)
- Ollama: The easiest way to run Gemma locally. Once you download the free Ollama Tool on Mac, Windows, or Linux, open your terminal and pull down a model weight variant:
ollama run gemma4:e4b - AI Hubs: Download raw model weights, fine-tune adapters, or explore community variations hosted on Kaggle Models or Hugging Face. [1, 7, 21, 22, 23]
Would you like to know how to locally install a specific Gemma model using Ollama, or are you interested in fine-tuning it with your own data?
[9] https://developers.google.com
[12] https://medium.com
[17] https://martini.art
[19] https://c3.unu.edu
[20] https://discuss.ai.google.dev
[22] https://eraoftech.medium.com