🖥️ Run AI Offline — Free & Private

Set up local LLMs with Ollama, OpenClaw, LM Studio, and more. No API keys. No cloud. No monthly bills.

⚠️ Before You Install

System Access: OpenClaw can read files, run terminal commands, and make changes to your computer.
Permissions: You'll be asked to grant various permissions during setup. Review each one carefully.
Local Models: Your data stays on your computer when using local models. No internet required after setup.
Cloud Models: If you connect cloud APIs (GPT, Claude, etc.), data is sent to those providers. Local-only is the default.
Your Responsibility: You are responsible for what OpenClaw does on your system. Review actions before confirming.

What You Can Do With Local AI

💬

Chat with AI

Talk to local models offline. No internet, no API keys, no usage limits. Completely private.

📁

Read & Write Files

Analyze documents, edit code, organize folders — all locally, all with natural language.

🌐

Code Generation

Qwen3-Coder and DeepSeek R1 run locally. Production-quality code, 80-150 tokens/second.

⚙️

Run Commands

OpenClaw executes terminal commands. Automate tasks, build scripts, manage your system.

🧠

Reasoning

DeepSeek R1 runs chain-of-thought reasoning locally. Think step-by-step, on your hardware.

🔌

Integrations

Connect to iMessage, Telegram, Discord, WhatsApp. Use AI everywhere, powered by local models.

💻 Hardware Requirements (April 2026)

8GB RAM — Entry Level

8GB RAM (Apple Silicon or 8GB VRAM GPU)
macOS 13+, Windows 10, or Linux
10GB free disk space
Models: Qwen3-Coder 8B, Llama 3.2 3B, Phi-4 Mini
~20-40 tokens/second

16GB RAM — Sweet Spot

16GB unified RAM or 12GB VRAM
Apple Silicon Mac or RTX 3060+
20GB free disk space
Models: DeepSeek R1 Distill 14B, Qwen 2.5 Coder 14B
~40-80 tokens/second

24GB+ RAM — Power User

24GB+ RAM (M-series Mac or 24GB VRAM)
Apple M1/M2/M3/M4 Pro/Max or RTX 4090
50GB+ free disk space
Models: Llama 4 Scout 17B, QwQ 32B, Mistral Small 24B
~60-150 tokens/second

🍎 Apple Silicon note: M-series Macs use unified memory — the GPU shares system RAM. An 8GB M2 Mac mini runs Qwen3-Coder 8B at ~40 tok/s. A 24GB M4 Pro runs Llama 4 Scout at ~60 tok/s. No discrete GPU needed.

📥 Installation (Ollama + OpenClaw)

🍎 macOS

🪟 Windows

🐧 Linux

Download OpenClaw

Download the latest macOS release from the official website:

Download from openclaw.ai GitHub Releases

Install the App

Open the downloaded .dmg file
Drag OpenClaw to your Applications folder
Open OpenClaw from Applications
If you see "unidentified developer" warning:
- Go to System Preferences → Privacy & Security
- Click "Open Anyway" next to the security warning

Grant Permissions

OpenClaw will request several permissions:

Accessibility: Required for keyboard shortcuts and automation
Full Disk Access: Required to read/write files in your folders
Automation: Required to control other apps

Go to System Preferences → Privacy & Security and enable these for OpenClaw.

Install Ollama (for local models)

OpenClaw uses Ollama for local AI models:

Download Ollama from ollama.com
Install it like any Mac app
Open Terminal and run: ollama pull qwen3-coder (best local coding model)

Recommended first models (April 2026):

ollama pull qwen3-coder — 8B, best coding model under 10B, 8GB RAM
ollama pull deepseek-r1:14b — reasoning model, 16GB RAM
ollama pull llama3.2 — general chat, 8GB RAM
ollama pull phi4-mini — ultra-fast, 8GB RAM

OpenClaw detects Ollama automatically. See AI Pricing Guide for the full model comparison.

Configure & Start

Open OpenClaw
Choose your default model (local or cloud)
Set up memory and preferences
Try asking: "What can you help me with?"

Download OpenClaw

Download the Windows installer:

Download from openclaw.ai GitHub Releases

Run the Installer

Run the downloaded .exe file
Follow the installation wizard
Choose install location (default is fine)
Launch OpenClaw from the Start Menu

Windows may show a "Windows protected your PC" warning. Click "More info" then "Run anyway" if you downloaded from the official source.

Grant Permissions

OpenClaw may request:

File system access: To read and write files
Network access: To connect to AI services
Windows Terminal: For running commands

Approve these in the Windows security dialogs that appear.

Install Ollama (for local models)

Download Ollama from ollama.com
Run the installer
Open PowerShell and run: ollama pull qwen3-coder

Configure & Start

Open OpenClaw from Start Menu
Choose your default model
Set up preferences
Start chatting!

Download OpenClaw

Download the Linux package (AppImage or .deb/.rpm):

Download from openclaw.ai GitHub Releases

Install

For AppImage:

chmod +x openclaw-*.AppImage
./openclaw-*.AppImage

For Debian/Ubuntu (.deb):

sudo dpkg -i openclaw-*.deb
sudo apt-get install -f  # Fix dependencies

For Fedora/RHEL (.rpm):

sudo rpm -i openclaw-*.rpm

Install Ollama

curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen3-coder

Run OpenClaw

Launch from your application menu or run:

openclaw

🏆 Best Local Models (April 2026)

Qwen3-Coder 8B ⭐

Best for: Coding, autocomplete
RAM: 8GB
Speed: 80-150 tok/s
Pull: ollama pull qwen3-coder
92 programming languages. Production frontend code.

DeepSeek R1 14B

Best for: Reasoning, analysis
RAM: 16GB
Speed: 40-55 tok/s
Pull: ollama pull deepseek-r1:14b
Chain-of-thought on budget hardware.

Llama 3.2 3B

Best for: General chat, quick tasks
RAM: 8GB
Speed: 40-60 tok/s
Pull: ollama pull llama3.2
Fast, reliable, well-tested.

Llama 4 Scout 17B

Best for: Massive context, multimodal
RAM: 24GB
Speed: 30-60 tok/s
Pull: ollama pull llama4-scout
10M context. Open weight. Multimodal.

💡 Tip: Install 2-3 models for different tasks. Qwen3-Coder for code, DeepSeek R1 for reasoning, Llama 3.2 for quick chat. Switch models in OpenClaw with one click. See the AI Pricing Guide for full comparisons including API models.

🛠️ Other Ways to Run Local AI

LM Studio

GUI app for downloading and chatting with models
Browse HuggingFace models inside the app
Easiest for beginners — no terminal needed
lmstudio.ai

OpenClaw + Ollama

AI assistant with memory, tools, and integrations
Telegram, Discord, iMessage, WhatsApp
File access, terminal, web search, automation
openclaw.ai

Ollama (CLI Only)

Terminal-based model runner
Lightweight, no GUI needed
Also exposes OpenAI-compatible API on port 11434
ollama.com

vLLM / SGLang

Production inference servers
For serving models to multiple users
Requires NVIDIA GPU or ROCm
Best for teams and self-hosted APIs

⚙️ Initial Configuration

After installation, OpenClaw will guide you through setup. Here's what to configure:

Model Selection

Local (Ollama): Free, private, offline-capable. Try qwen3-coder for coding or deepseek-r1:14b for reasoning.
Cloud: More powerful but sends data to servers. Add API keys for ChatGPT, Claude, etc.

Memory Setup

OpenClaw can remember information across conversations. In settings:

Enable "Long-term memory" to remember facts about you
Set memory location (default is fine)
Configure what to remember/forget

Integrations (Optional)

Connect to messaging platforms in Settings → Integrations:

iMessage: Chat with AI via Messages app (Mac only)
Telegram: Create a bot for Telegram chats
Discord: Add AI to your Discord server

🔧 Troubleshooting

"Ollama not found" error

Make sure Ollama is installed and running:

# Check if Ollama is running
ollama list

# If not, start it
ollama serve

"Permission denied" errors

On macOS, make sure OpenClaw has:

Accessibility access (System Preferences → Privacy & Security)
Full Disk Access (System Preferences → Privacy & Security)

Slow responses

Use a smaller model (phi4-mini or llama3.2:3b for 8GB RAM)
Close other resource-intensive apps
Check you have enough RAM for the model size

Need more help?

Documentation GitHub Issues Discord Community