How to Run OpenClaw with Local LLMs?

Running OpenClaw with local models means zero API costs, full privacy, and offline capability. All processing happens on your hardware — no data leaves your machine.

Two main tools exist for local models: Ollama (CLI-based, best for servers) and LM Studio (GUI-based, best for desktops). Both provide OpenAI-compatible APIs that OpenClaw connects to seamlessly.

Recommended local models for OpenClaw: - Llama 3.1 8B: Best general-purpose, 8 GB RAM - Qwen 2.5 14B: Strong coding and reasoning, 16 GB RAM - Mistral Nemo 12B: Good multilingual support, 16 GB RAM - Phi-3 3.8B: Ultra-lightweight, 4 GB RAM, suitable for Raspberry Pi

Hardware guide: Apple Silicon Macs are the sweet spot for local models — unified memory lets the GPU access all system RAM. An M2 Mac Mini with 16 GB runs 8B models at 20-30 tokens/second. For x86, a gaming GPU (RTX 3060 12GB+) provides the best performance.

The quality-cost trade-off is real. Local 8B models handle simple tasks well but struggle with complex reasoning, long context, and nuanced instructions compared to Claude Sonnet or GPT-4o. Consider a hybrid approach: local model for simple daily queries, cloud model for important tasks.

Tip: Use OpenClaw's workspace feature to create one workspace with a local model (free daily driver) and another with Claude (for quality-critical tasks).

# Ollama approach
ollama pull llama3.1
openclaw config set provider ollama
openclaw config set model llama3.1

# LM Studio approach
# Start LM Studio, load a model
openclaw config set provider openai-compatible
openclaw config set baseUrl http://localhost:1234/v1