# Gemma 4 AI

> Practical, tested guides for running Google's Gemma 4 models locally — hardware requirements, quantization, deployment with Ollama and vLLM, fine-tuning, and troubleshooting. Numbers come from real local runs, not marketing claims.

## Core Guides
- [Gemma 4 Hardware Requirements (8GB/16GB/32GB)](https://gemma4-ai.com/blog/gemma4-hardware/): Exact RAM/VRAM for E2B, E4B, 26B MoE, 31B Dense across Q4/Q8/FP16, KV-cache math, and device-by-device guidance.
- [Gemma 4 Download Guide](https://gemma4-ai.com/blog/gemma4-download-guide): Where and how to get the weights.
- [Which Gemma 4 Model Should I Use?](https://gemma4-ai.com/blog/gemma4-which-model): Choosing between E2B, E4B, 26B MoE and 31B Dense.
- [Run Gemma 4 with Ollama](https://gemma4-ai.com/blog/how-to-run-gemma4-with-ollama): Step-by-step local deployment.
- [Gemma 4 26B vs 31B](https://gemma4-ai.com/blog/gemma4-26b-vs-31b): Head-to-head comparison.
- [Gemma 4 Speed Optimization](https://gemma4-ai.com/blog/gemma4-speed-optimization): Squeezing more tokens/sec from your hardware.

## Key Facts (Gemma 4, current as of 2026-05)
- Models: E2B (2B), E4B (4B), 26B MoE, 31B Dense.
- Minimum to start: E2B at Q4 needs ~1.5GB; runs on an 8GB machine.
- 31B Dense at Q4 needs ~20GB VRAM; FP16 needs ~62GB; KV cache at full 262K context adds ~22GB.
- Recommended quantization default: Q4_K_M (~75% size, 95–97% quality).

## About
- Authoritative sources we build on: Google AI Studio (aistudio.google.com), Hugging Face (huggingface.co), Ollama (ollama.com), official model card (ai.google.dev).
- Contact / canonical site: https://gemma4-ai.com