Gemma 4 Local

Run Google's Gemma 4 AI entirely on your Mac. No internet. No subscriptions. Free forever.

---

What is this?

A native Mac app that lets you chat with Google's Gemma 4 AI model — completely offline. Your conversations never leave your machine. No account needed, no API keys, no monthly fees.

Features

• 100% offline — works without internet, on a plane, anywhere
• Private — zero telemetry, no data sent anywhere, ever
• Thinking mode — toggle with Cmd+T for step-by-step reasoning
• Conversation history — auto-saved, searchable across all chats
• Inline controls — temperature, tokens, thinking mode right above the input
• Multimodal — text, images and audio (Gemma 4 E4B)
• Benchmark slideshow — see how it compares to paid models
• Signed & notarized — Apple Developer ID, opens without security warnings

Requirements

• Mac with Apple Silicon (M1, M2, M3, M4)
• 16 GB RAM
• macOS 13 Ventura or later
• ~7 GB free disk space

Install

1. Download the DMG
2. Open it, drag to Applications
3. Open the app, click "Iniciar modelo"
4. Wait ~10 seconds while the model loads into memory
5. Chat

The first launch downloads the model (~5 GB) from HuggingFace. After that, everything runs locally.

_{Works without internet. On a desert island, on a plane, at your grandparents' village. If your Mac turns on, your AI works.}

Benchmarks

Gemma 4 E4B (free, local) vs paid API models:

Benchmark	Gemma 4 E4B	GPT-4o mini	Claude 3.5 Haiku
MMLU Pro	69.4	63.1	65.0
LiveCodeBench	52.0	23.4	31.4
GPQA Diamond	58.6	44.2	41.6

It beats every paid lightweight model. See the full comparison →

_{$0 per month. Forever. No subscriptions, no credit card, no "your free trial has expired".}

How it works

The app bundles a Python virtual environment with vMLX (an MLX-based inference engine optimized for Apple Silicon). When you click "Iniciar modelo", it starts a local server and loads the Gemma 4 E4B model into GPU memory. The Electron frontend communicates with it via a local API. Nothing touches the network.

Keyboard shortcuts

Shortcut	Action
Cmd+T	Toggle thinking mode
Cmd+N	New conversation
Enter	Send message
Shift+Enter	New line
Cmd+Q	Quit (stops model, frees RAM)

_{Your data stays in. Not Google, not Facebook, not anyone. A bubble only you control.}

Security audit

The vMLX engine was fully audited before integration:

• No telemetry — zero outbound connections
• No eval/exec — only mx.eval() (MLX GPU sync)
• No pickle — all weights loaded via safetensors
• API key auth — optional, for network exposure
• Agentic tools disabled — shell execution features are off by default

Uninstall

# Remove the app
rm -rf "/Applications/Gemma 4 Local.app"

# Remove the model (~5 GB)
rm -rf ~/.cache/huggingface/hub/models--mlx-community--gemma-4-e4b-it-4bit

# Remove the project (if installed from source)
rm -rf ~/.gemma4-local

Nothing else is modified. No daemons, no PATH changes, no config files.

_{Anywhere. Anytime. Even in space. Well, you might need oxygen before AI up there.}

Legal

This is an independent project. Not affiliated with Google, Alphabet, or DeepMind. "Gemma" is a trademark of Google LLC. Model weights are licensed under Apache 2.0 by Google. See NOTICE for full trademark attributions.

License

MIT — the app code is yours to use, modify, and distribute.

---

_{Built by Mario Hernandez · @mariohernandez}