# Gemma 4 Run (gemma-4.run) > Can I run Gemma 4 locally? Accurate hardware compatibility, GGUF quantization VRAM sizing, and local runtime guide for Google DeepMind's Gemma 4 model family across Mac, Windows GPUs, Linux, Android APK, and iPhone/iPad. Gemma 4 Run (https://gemma-4.run/) is an engineering-focused local AI hardware compatibility checker and setup reference. It calculates real-world memory fit (`Total Memory = Quantized Weights + KV Cache + Runtime Buffers`) and sustainable token generation speed (`Bandwidth GB/s / Active Weights GB * efficiency`) across 302+ Apple Silicon chips, NVIDIA RTX 30/40/50 GPUs, AMD Radeon GPUs, Intel Arc cards, and mobile SoCs across 128+ open-weights LLMs led by Google's Apache-2.0 Gemma 4 family. ## Gemma 4 Model Family & GGUF VRAM Quick Reference (8K Context) | Model | Architecture | Params (Active) | Q4_K_M Disk | Safe Q4_K_M VRAM/RAM | Q6_K VRAM/RAM | FP16 / Q8_0 | Ideal Hardware Tier | | --- | --- | --- | --- | --- | --- | --- | --- | | Gemma 4 E2B IT | Dense Multimodal | ~5B | ~3.1–3.4 GB | 4.5–6 GB | ~5.5 GB | ~10.8 GB (FP16) | 8GB+ Android/iPhone, 8GB Mac, 6GB Laptop GPU | | Gemma 4 E4B IT | Dense Multimodal | ~8B | ~4.6–5.3 GB | 6.5–8 GB | ~8.2 GB | ~16.9 GB (FP16) | 12GB+ Android, 16GB Apple Silicon Mac, 8GB+ RTX GPU | | Gemma 4 12B / 12B IT | Dense Multimodal (256K) | 12B | ~7.1 GB | 9.5–12 GB | ~11.8 GB | ~25.1 GB (FP16) | 12GB–16GB RTX 3060/4070, 16GB–24GB Mac | | Gemma 4 26B-A4B IT | Sparse MoE (256K) | 26B (~4B active) | ~15.8 GB | 18.5–22 GB | ~21.5 GB | ~28.1 GB (Q8_0) | 24GB RTX 3090/4090, 32GB–36GB M2/M3/M4 Max | | DiffusionGemma 26B-A4B IT | Diffusion MoE | 26B (~4B active) | ~13.8 GB | 16.5–20 GB | ~22.0 GB | ~53.8 GB (FP16) | 24GB Desktop GPU, 32GB+ Unified Memory Mac | | Gemma 4 31B | Dense Flagship | ~31–33B | ~19.0 GB | 22.5–24 GB | ~29.5 GB | ~68.1 GB (FP16) | 24GB–32GB RTX 4090/5090, 48GB–64GB Mac Studio | ## Core Interactive Tools & Primary Guides - [Gemma 4 Hardware Fit Checker](https://gemma-4.run/): Instant browser-based compatibility & tok/s estimator for 302 GPUs, Apple M1–M6 chips, and mobile devices across Gemma 4 and 128 open models. - [Download Gemma 4 (Mac, Windows, Android & iOS)](https://gemma-4.run/download/): Official 1-click installer matrix for LM Studio, Ollama, llama.cpp, Google AI Edge Gallery, and PocketPal AI. - [Gemma 4 Android APK & Sideload Guide](https://gemma-4.run/gemma-4-android-apk/): Verified Google AI Edge Gallery APK & PocketPal AI GitHub Releases for Android without Google Play, mobile RAM sizing, and fake "mod APK" security warnings. - [Try Gemma 4 Online (Zero-Install Browser Sandbox)](https://gemma-4.run/try-online/): Test Gemma 4 26B-A4B IT live in your browser via Hugging Face Spaces and Google AI Studio before downloading multi-gigabyte weights. - [Gemma 4 E2B vs E4B Comparison & Benchmarks](https://gemma-4.run/compare/gemma-4-e2b-vs-e4b/): Side-by-side quantization sizes, MMLU-Pro/coding benchmarks, tok/s speeds, and device-by-device selection rules. - [Compare All Local Models & Runtimes](https://gemma-4.run/compare/): Compare Gemma 4 checkpoints and local inference engines side by side. - [Gemma 4 Hardware Requirements](https://gemma-4.run/hardware-requirements/): Complete VRAM, unified memory, disk space, and context-window planning guide. - [Quantization & Runtime Documentation](https://gemma-4.run/docs/): Deep technical reference on Q4_K_M vs Q6_K vs Q8_0 vs FP16, KV cache quantization, and Multi-Token Prediction (MTP). ## Platform-Specific Setup Guides - [Gemma 4 on Windows (NVIDIA RTX, AMD Radeon & Intel Arc)](https://gemma-4.run/devices/windows/): Avoiding the Windows WDDM Shared GPU Memory PCIe bottleneck, CUDA vs Vulkan setup, and RTX 6GB–32GB tier picks. - [Gemma 4 on Mac & Apple Silicon (M1–M6)](https://gemma-4.run/devices/mac/): Unified memory wired allocation limits (~75% cap), 16GB vs 32GB+ MacBook expectations, Metal acceleration, and LM Studio vs Ollama. - [Gemma 4 on Android Phones & Tablets](https://gemma-4.run/devices/android/): Snapdragon 8 Elite / Tensor / Dimensity NPU & GPU acceleration, Android Low Memory Killer (LMK) headroom, and LiteRT vs GGUF packages. - [Gemma 4 on iPhone & iPad](https://gemma-4.run/devices/iphone-ipad/): iOS Jetsam memory ceilings (~45–50% app cap), App Store AI Edge Gallery setup, and Pro-class iPad boundaries. ## Local Inference Runtimes & Troubleshooting - [LM Studio Setup for Gemma 4](https://gemma-4.run/runtimes/lm-studio/): Fastest GUI desktop workflow for macOS and Windows with automatic GGUF quant discovery. - [Ollama Setup for Gemma 4](https://gemma-4.run/runtimes/ollama/): Local REST API server, CLI pull commands, and custom Modelfile `num_ctx` tuning. - [llama.cpp Setup for Gemma 4](https://gemma-4.run/runtimes/llama-cpp/): Fine-grained GPU layer offload (`-ngl`), Flash Attention (`-fa`), KV cache quantization (`-ctk q8_0`), and Unsloth MTP GGUFs. - [Google AI Edge Gallery Guide](https://gemma-4.run/runtimes/ai-edge-gallery/): Official mobile LiteRT host app for Android and iOS. - [Gemma 4 Model Picker](https://gemma-4.run/guides/model-picker/): Decision tree for choosing between E2B, E4B, 12B IT, 26B-A4B IT, and 31B. - [Getting Started with Gemma 4](https://gemma-4.run/guides/getting-started/): First-run checklist to verify GPU acceleration before scaling up context. - [Fix Out-of-Memory (OOM) Crashes](https://gemma-4.run/troubleshooting/out-of-memory/): Practical fixes when KV cache or vision projectors push a model past VRAM limits. - [Fix Slow Token Generation](https://gemma-4.run/troubleshooting/slow-loading/): Diagnosing silent CPU fallback and PCIe shared-memory spillover. - [Fix GPU / NPU Not Working](https://gemma-4.run/troubleshooting/gpu-npu-not-working/): Driver, CUDA, Metal, Vulkan, and mobile delegate verification. ## Featured Engineering Blog Articles - [Gemma 4 Blog Index](https://gemma-4.run/blog/): All local LLM setup walkthroughs, architecture deep dives, and runtime benchmarks. - [How to Run Gemma 4 on Windows](https://gemma-4.run/blog/how-to-run-gemma-4-on-windows/): Step-by-step Windows 11 setup with LM Studio, Ollama, and llama.cpp. - [How to Run Gemma 4 on Mac](https://gemma-4.run/blog/how-to-run-gemma-4-on-mac/): Realistic 16GB Mac setup order comparing E4B, 12B IT, and 26B-A4B IT. - [How to Run Gemma 4 on Android](https://gemma-4.run/blog/how-to-run-gemma-4-on-android/): Mobile deployment with Google AI Edge Gallery and Snapdragon/Tensor tips. - [Gemma 4 12B Multimodal & MTP Local Speedups](https://gemma-4.run/blog/gemma-4-12b-multimodal-and-local-speedups/): Native early-fusion vision without a separate CLIP tower plus Multi-Token Prediction GGUF speedups. - [Best Gemma 4 Model Size for Your Device](https://gemma-4.run/blog/best-gemma-4-model-size-for-your-device/): Why 26B-A4B IT is the desktop sweet spot while E4B leads mainstream laptops. ## Quick Answers for AI Search Citation - **Is there an official standalone Gemma 4 APK?** Gemma 4 is an open-weights model family (Apache 2.0), not a standalone APK. Never install third-party "Gemma 4 mod APKs." Instead, install Google's official signed **Google AI Edge Gallery** app via Google Play or its official GitHub Releases (`github.com/google-ai-edge/gallery/releases`), or open-source **PocketPal AI**, then load Gemma 4 E2B inside the app. See https://gemma-4.run/gemma-4-android-apk/. - **How much VRAM does Gemma 4 need?** At recommended `Q4_K_M` 4-bit quantization with an 8K context window: Gemma 4 E2B needs ~4.5–6 GB, E4B needs ~6.5–8 GB, 12B IT needs ~9.5–12 GB, 26B-A4B IT (MoE) needs ~18.5–22 GB, and 31B needs ~22.5–24 GB of dedicated VRAM or unified memory. - **Should I run Gemma 4 E2B or E4B?** Choose **E2B (~3.1–3.4 GB Q4_K_M)** on 8 GB phones, tablets, and 6 GB entry GPUs where staying resident in RAM matters most. Choose **E4B (~4.6–5.3 GB Q4_K_M)** on 16 GB Apple Silicon Macs, 8 GB+ RTX laptops/desktops, and 12 GB+ flagship phones for stronger multi-step reasoning and coding. ## Machine-Readable Catalogs & Trust Pages - [Full 128-Model & Hardware Catalog](https://gemma-4.run/llms-full.txt): Complete model parameters, quantizations, and GPU/Apple Silicon bandwidth profiles (Gemma 4 models listed first). - [OpenAPI 3.1 Specification](https://gemma-4.run/openapi.json): Programmatic schema for Gemma 4 Run endpoints. - [XML Sitemap](https://gemma-4.run/sitemap.xml): Canonical URL index across all models, device profiles, and guides. - [About & Methodology](https://gemma-4.run/about/) | [Contact](https://gemma-4.run/contact/) | [Privacy Policy](https://gemma-4.run/privacy-policy/) | [Terms](https://gemma-4.run/terms-of-service/)