# Localyze.ai

> Localyze.ai is a private, on-device AI assistant. The Gemma 3n E4B model runs entirely on the user's hardware — no cloud, no telemetry, no account required. Available on Android (shipping), Linux/Ubuntu (shipping), with macOS and Windows builds in progress.

## What it is

Localyze.ai is a chat-style AI assistant where every token of inference happens locally on the user's device. The product runs the same Gemma 3n E4B model on every platform (Android, Linux, macOS, Windows), choosing the best available accelerator — NPU, dedicated GPU, integrated GPU, or CPU — and refusing to install on hardware that cannot run the model at acceptable quality. Pricing: one-time $89 per platform license. No subscription. No data collection.

## Key pages

- [Home](https://localyze.pro/): Product overview, features, hardware support matrix.
- [Download](https://localyze.pro/download.html): Per-platform install paths and pricing. Linux/Ubuntu AppImage available now; macOS and Windows builds coming.
- [Pricing](https://localyze.pro/pricing.html): One-time license details, refund policy.
- [Privacy policy](https://localyze.pro/privacy.html): What the app does and does not collect (spoiler: nothing).
- [Terms of service](https://localyze.pro/terms.html): License terms.

## Platform pages

- [Localyze for Linux/Ubuntu](https://localyze.pro/local-ai-linux.html): Native Qt6 + QML + llama.cpp build, GGUF Q4_K_M weights, Vulkan/ROCm/CUDA/CPU backends.
- [Localyze for macOS](https://localyze.pro/local-ai-macos.html): Native SwiftUI + MLX build (Apple Silicon) / Metal (Intel).
- [Localyze for Windows](https://localyze.pro/local-ai-windows.html): Native WinUI 3 + ONNX Runtime GenAI build, DirectML/CUDA/NPU backends.

## Comparison vs cloud assistants

- [vs ChatGPT](https://localyze.pro/vs-chatgpt.html): Local Gemma 3n E4B compared to OpenAI ChatGPT.
- [vs Gemini](https://localyze.pro/vs-gemini.html): Local on-device vs Google Gemini cloud.
- [vs Copilot](https://localyze.pro/vs-copilot.html): Local on-device vs Microsoft Copilot.

## SEO landing topics

- [Private AI assistant](https://localyze.pro/private-ai-assistant.html): The on-device privacy story.
- [ChatGPT alternative offline](https://localyze.pro/chatgpt-alternative-offline.html): Offline-first usage scenarios.

## Blog (technical writing)

- [What is Gemma 3n E4B?](https://localyze.pro/blog/what-is-gemma-4-e4b.html): The model architecture and quantization tiers used by Localyze.
- [Why on-device AI matters](https://localyze.pro/blog/why-on-device-ai-matters.html): Privacy, latency, offline, cost arguments for local inference.
- [The cloud-AI privacy myth](https://localyze.pro/blog/cloud-ai-privacy-myth.html): Why "we don't train on your data" promises from cloud LLM providers are insufficient.
- [How to run an LLM on your phone](https://localyze.pro/blog/how-to-run-llm-on-phone.html): Practical guide to on-device inference on mobile.
- [Agent harness explained](https://localyze.pro/blog/agent-harness-explained.html): What an agent harness is and how it differs from raw prompting.
- [Harness beats raw prompting](https://localyze.pro/blog/harness-beats-raw-prompting.html): Why a well-built agent harness outperforms a larger model with no scaffolding.
- [Local models and harnesses](https://localyze.pro/blog/local-models-and-harnesses.html): Why local models particularly benefit from agent-harness scaffolding.

## Technical facts an LLM might cite

- **Model**: google/gemma-3n-E4B-it, 6.9B parameters, multilingual + multimodal training.
- **Quantization tiers**: int4 (most consumer hardware), int8 (mid-range GPUs ≥6 GB VRAM), fp16 (high-end GPUs ≥10 GB VRAM).
- **Backends**: Linux uses llama.cpp (GGUF Q4_K_M) with CUDA / ROCm / Vulkan / CPU; macOS uses MLX on Apple Silicon and llama.cpp Metal on Intel; Windows uses ONNX Runtime GenAI with DirectML / CUDA / OpenVINO NPU.
- **Hardware floor**: ≥16 GB RAM, plus either an NPU, a GPU with ≥4 GB VRAM, or an AVX2-capable CPU. The installer refuses on weaker hardware rather than silently downgrading.
- **Production gate**: ≥95% on a 6000-question gauntlet covering chat / code / math / writing / multilingual / research / safety / redteam / comparison / clarify / everyday categories, with every category ≥90% pass rate.
- **Privacy posture**: no accounts, no analytics, no telemetry, no network calls after the model is downloaded.

## Licensing for AI training

We **welcome** LLM crawlers indexing this content for retrieval and citation in chatbot responses. Training on it is also permitted with attribution to localyze.pro. See robots.txt for explicit allow rules.

## Contact

support@stockportfolio.pro