About Gemma 4 Run
Editorial mission, memory calculation methodology, source verification standards, and independence disclosures for Gemma 4 Run.
Last updated: 2026-11-15
Editorial Mission and Independence
Gemma 4 Run (https://gemma-4.run) is an independent technical resource and hardware sizing platform built for developers, ML engineers, students, and hobbyists running open-weight AI models locally. Our goal is straightforward: help you know whether a specific checkpoint fits your laptop, desktop GPU, Apple Silicon Mac, smartphone, or single-board computer before you download tens of gigabytes of weights.
Gemma 4 Run is a community-maintained website and is not affiliated with, endorsed by, or sponsored by Google LLC or Google DeepMind. "Gemma" is a trademark of Google LLC. For official model weights, evaluation cards, and Apache 2.0 or custom license agreements, always consult the primary publisher repositories such as the official Google Gemma 4 model collection on Hugging Face.
How Our Memory and Speed Estimates Are Calculated
Local model inference depends on far more than raw parameter count. Our compatibility engine combines quantization footprint modeling with hardware architecture rules across four distinct memory pools:
- Weight Footprint & Quantization Overhead: Starting from each checkpoint's parameter count or verified GGUF file size (
Q2_K,Q3_K_M,Q4_K_M,Q6_K,Q8_0,FP16), we estimate resident weight size plus runtime tensor overhead and ~0.5 GB baseline context allocation (llama.cpp/ CUDA / Metal runtime context). - Operating System Memory Ceilings:
- Discrete Desktop & Laptop GPUs (NVIDIA, AMD, Intel): Models comfortably fit when total resident memory stays below ~85% of dedicated VRAM. Above that margin, long context windows can spill layers over PCIe into system RAM, sharply lowering generation throughput.
- Apple Silicon Unified Memory: Because macOS shares physical LPDDR memory across CPU and GPU, our checker budgets roughly 75% of total RAM as practical wired GPU headroom unless custom
sysctllimits are configured. - Mobile Devices (Android & iOS): Android Low Memory Killer (LMK) and iOS Jetsam daemons terminate background-heavy processes automatically; our sizing rules budget roughly 50–55% of physical RAM on phones and tablets.
- Bandwidth-Limited Token Throughput: Autoregressive decoding speed (
tok/s) is governed primarily by effective memory bandwidth (GB/s) divided by active weight size per token, adjusted for Metal, CUDA, mobile, or PCIe offload harmonic penalties. For sparse Mixture-of-Experts (MoE) checkpoints such as Gemma 4 26B-A4B, total VRAM must hold all 26B parameters while throughput scales against the ~4B active parameters per forward pass.
Treat every browser estimate as a practical planning baseline rather than a hard guarantee. Context length (4K vs 128K), vision/audio encoders, flash attention settings, driver versions, and background applications all shift real-world memory consumption.
Editorial Standards and Source Verification
Every guide, troubleshooting playbook, and model compatibility card on Gemma 4 Run is structured around verifiable technical documentation:
- Model Metadata & File Sizes: Cross-checked against official Hugging Face model cards, GGUF repository manifests, and publisher release notes.
- Hardware Profiles: Derived from official silicon and board specifications published by NVIDIA, AMD, Intel, Apple, Qualcomm, and Raspberry Pi.
- Runtime Workflows: Tested and documented against upstream documentation for
llama.cpp, Ollama, LM Studio, ComfyUI, PocketPal, and Google AI Edge Gallery. - Open-Source Lineage: Portions of the browser WebGPU hardware detection logic and initial model/device catalog draw upon the open-source canirun.ai project and its contributors, extended here with Gemma 4 sizing matrices, original device guides, and operational troubleshooting playbooks.
We publish clear disclaimers whenever an estimate is modeled rather than benchmarked on physical hardware, and we never fabricate download links or claim unsupported third-party APK mirrors.
Corrections, Feedback, and Machine-Readable Endpoints
Model quantization tables, runtime flags, and hardware drivers evolve rapidly. If you spot an outdated GGUF size, incorrect VRAM figure, broken reference link, or runtime regression, please reach out via our Contact & Support page with the page URL and primary documentation link so our editorial team can verify and update the record.
Developers and automated tooling can also inspect our published documentation index via LLMs Manifest and OpenAPI Specification, or review our data practices in the Privacy Policy and Terms of Service.