Contents

• Start Here
• Models
  • Core Models
  • Variants
• Inference
  • Local
  • Hosted
• Fine-Tune
• Tutorials
• Demos and Applications
• Gemma 4 Good Challenge
• Gemma in Space
• Research and Evaluation

Start Here

• Gemma Documentation — Official documentation for selecting, running, tuning, and deploying Gemma models.
• Get Started with Gemma — Get started running inference with the multimodal Gemma 4 models.
• Gemma Cookbook — Maintained notebooks, examples, workshops, and end-to-end applications.
• Gemma Skills — Reusable Agent Skills for selecting, running, and training Gemma models.
• Gemma Events — Overview of upcoming Gemma events.
• Gemma on X — For news, announcements, and updates about Gemma.

Models

Core Models

• Gemma 4 Overview — An overview of the capabilities, architecture, and more of Gemma 4 models.
• Gemma 4 Model Card — Architecture, training, evaluation, safety, and usage details.
• Gemma 4 (Hugging Face) — Gemma 4 checkpoints (+ assistant models) on Hugging Face.
• Gemma 4 QAT — Quantization-aware checkpoints for local and server inference.
• Gemma 4 Mobile QAT — Mobile-optimized checkpoints for the E2B and E4B models.
• Previous Gemma Model Cards — Model cards for Gemma 1–3 and earlier Gemma families.

Variants

• DiffusionGemma — Experimental discrete-diffusion text generation based on Gemma 4.
• EmbeddingGemma — Compact embedding model designed for retrieval and on-device use.
• FunctionGemma — Foundation for building specialized function-calling models.
• MedGemma — Models optimized for medical text and image comprehension.
• PaliGemma 2 — Vision-language models for detailed image understanding tasks.
• ShieldGemma 2 — Image-safety classifier built on Gemma 3.
• T5Gemma 2 — Encoder-decoder models for contextual understanding and generation.
• TranslateGemma — Translation models covering 55 languages.
• TxGemma — Models for therapeutic-development research.
• VaultGemma — Language model trained with differential privacy.
• DataGemma — Models and recipes for grounding responses with Data Commons.
• RecurrentGemma — Open models based on the recurrent Griffin architecture.
• Gemma Scope 2 — Open sparse autoencoders and interpretability tooling for studying Gemma 3.
• Gemma-APS — Abstractive proposition segmentation for decomposing text into meaningful claims.
• Cell2Sentence-Scale — A Gemma 2 27B model fine-tuned for single-cell biology.
• DolphinGemma — Uses dolphin audio to help scientists study how dolphins communicate.

Inference

Local

• HF Transformers — Python library for loading, running, and fine-tuning Hugging Face models.
• llama.cpp — LLM inference in C/C++ with GGUF quantization.
• Unsloth — Local UI to run and train LLMs and diffusion models.
• Ollama — Get up and running with large language models locally.
• LM Studio — Desktop application to discover, download, and run local models.
• vLLM — High-throughput and memory-efficient LLM serving engine.
• SGLang — Fast serving framework for large language models and vision-language models.
• AI Edge Gallery — On-device ML models and examples for mobile and edge devices.
• LiteRT — Google's runtime for on-device ML deployment.
• JAX — Official Gemma reference implementation in JAX and Flax.
• React Native — Run on-device Gemma models within React Native using ExecuTorch.
• GenieX — Run Gemma on Qualcomm hardware.
• Docker — Run Gemma 4 in Docker.

Hosted

• Gemini Enterprise Agent Platform (Formerly Vertex AI) — Fully managed enterprise AI platform on Google Cloud.
• OpenRouter — Unified API routing to multiple AI model providers.
• Cerebras — High-speed Gemma 4 inference on Cerebras.
• NVIDIA — Optimized TensorRT-LLM and NVFP4 checkpoints.
• AMD — Support for AMD ROCm GPUs and processors.
• AI Studio — Web-based prototyping and development environment.
• Cloud Run — Deploy containerized Gemma services with autoscaling GPUs.
• LiveKit — Real-time multimodal voice and video inference infrastructure.
• Together AI — Cloud platform for running and fine-tuning open source models.
• Modal — Run and deploy Gemma 4 on the Modal platform.
• Fireworks — Run and deploy Gemma 4 on the Fireworks.AI platform.
• BaseTen — Run and deploy Gemma 4 on the BaseTen platform.
• Runpod — Experiment, train, fine-tune, and deploy Gemma.
• Cloudflare — Run Gemma 4 on the Workers AI LLM Playground.

Fine-Tune

• Fine-Tune Gemma — Official framework guide covering Keras, JAX, Hugging Face, Unsloth, Axolotl, and Google Cloud.
• Gemma Cookbook: Training — Official fine-tuning notebooks and training recipes.
• Tunix — JAX-native library for post-training generative models.
• Unsloth Gemma 4 fine-tuning guide — Train Gemma 4 E2B, E4B, 12B, 26B A4B and 31B with Unsloth.
• Gemma Multimodal Tuner — Fine-tune Gemma 3n and Gemma 4 with text, images, and audio on Apple Silicon.
• MLX Tune — MLX-native SFT, preference tuning, and multimodal fine-tuning with Gemma 4 support.

Tutorials

• A Visual Guide to Gemma 4
• A Visual Guide to Gemma 4 12B
• A Visual Guide to DiffusionGemma
• A Visual Guide to the Gemma 4 Drafters
• Variable Aspect Ratio and Variable Resolutions in Gemma 4
• How to Use Transformers.js in a Chrome Extension
• How to run a local coding agent with Gemma 4 and Pi
• How to run Gemma 4 with OpenClaw
• How a Small Fix Improves Gemma 4 Vision Performance
• While I slept, my 5-year-old MacBook ran Gemma 4 locally and indexed a year of video
• Fine-tuning Gemma 4 12B on your own data
• Turning Gemma 4 into an Old Korean Translator

Demos and Applications

• Gemma 4 Vision Token Budget — Explore the effect of image resolution and visual-token budgets.
• Concurrent Gemma — Run and compare multiple concurrent local Gemma instances.
• See what 3 builders are making with Gemma 4 — Various applications developed by the community.
• AIventure — A 2D grid-based adventure game built with Phaser 3 and Angular with Gemma driving it.
• Gemma Chat — Local AI chat + coding agent for Apple Silicon, powered by Gemma 4 via MLX / Supports Ollama.
• Build with Gemma 4 and Haystack — Runnable notebook covering RAG, visual question answering, a multimodal weather agent, and GitHub tool discovery.
• Gemma 4 Browser Extension — Local browser agent powered by Gemma 4, WebGPU, and Transformers.js.
• WebGemma — Browser playground and interactive model timeline powered by WebGPU and Transformers.js.
• Controlling an iOS simulator — Gemma 4 using Argent to control an iOS simulator showcasing its capabilities in agentic workflows.
• Automated Video Segmentation & Tracking — A demo that uses Gemma 4 + Falcon Perception for video tracking.
• Parking Lot Car Detection & Segmentation — Gemma 4 analyzes the scene, decides the questions, generates prompts, and calls SAM 3.1 as a tool. SAM 3.1 segments and returns results.
• Gemma 4 and MTP as a Marathon Engine — Benchmarks speculative decoding across increasing context lengths.
• Cactus Hybrid — Post-trained Gemma 4 models to recognize when they are wrong, run on any framework.
• Damage Scout — Damage Scout samples frames from a rental car walkaround, sends them to Gemma 4, gets back structured findings and box coordinates, then renders an annotated damage report in under 6 seconds.
• MedGemma Impact Challenge — The winners of the MedGemma hackathon to build human-centered AI applications with MedGemma.
• Gemma-Translator — A fully offline device powered by Gemma 4 E2B built with Google Antigravity.
• Real-Time Voice AI with Gemma 4 — Open-source cascaded voice stack using Gemma 4 for low-latency reasoning.

Gemma 4 Good Challenge

Amazing projects that harness the power of Gemma 4 to drive positive change and global impact.

• Trido — A Voice-Driven AI Whiteboard Built for the Teacher Nobody Builds For.
• CodeBuddy — AI Python Tutor for Indonesian Students.
• Port-a-Prof — Deeper learning, wherever you are.
• TriageMate — Offline-first Clinical AI for Ghana's Community Health Officers.
• ORCA-G4 — On-device oral cancer intelligence for 900,000 ASHA workers in rural India.
• DEMENTOR — Edge AI Triage for Dementia Care.
• PreVillage — A source-backed navigator for Nepal’s government services, built to find the office route, not just the form.
• BrailleOut — An assistive device that reads the text and images from real-world and converts it to Braille using Gemma 4 and Ollama.
• Gem-Care — Gemma-4-Enriched with Multimodal Clinical-context Adaptation for Recognition Enhancement of Non-Normative Speech.
• Trajectix — An Agentic Flight Recorder for AI Infrastructure Safety.
• TrueVoice — AI Voice Deepfake Detector.
• AI Conceptualizer — 3D visualizations for mechanistic interpretability and "concept spectroscopy".
• Acuífero·Vigía — Hybrid edge-and-citizen flood early warning for Argentina's Litoral, where every minute of warning is a life.
• ResQ — Offline Multilingual Disaster Response Coach on Gemma 4 E2B.

Gemma in Space

• Starcloud-1 — Starcloud deployed and ran Gemma in orbit aboard an H100 GPU.
• NASA — NASA runs Gemma in orbit to analyze satellite imagery and compress visual data into text for rapid, low-bandwidth disaster response.

Research and Evaluation

• Gemma 4 Technical Report — The technical report covering Gemma 4 E2B, E4B, 12B, 26B A4B, and 31B.
• DiffusionGemma Technical Report — The technical report covering DiffusionGemma.
• Artificial Analysis — Intelligence, Performance & Price Analysis.
• ChessBench — Chess LLM Benchmark Leaderboard.
• TERMS-Bench — A benchmark for LLM negotiation agents based on economic negotiation.

Footnotes

This is not an officially supported Google product. This project is not eligible for the Google Open Source Software Vulnerability Rewards Program.