Best Local AI Models for OpenClaw
Recommended local models to run with OpenClaw.
How to Choose a Model
- • Task type — Coding, conversation, reasoning
- • Hardware — RAM, VRAM, GPU availability
- • Language — English, multilingual, coding-only
- • Speed — Smaller = faster, larger = smarter
Recommended Models
Llama 3.2
Best OverallMeta's latest. Great reasoning, excellent English, available in 1B-90B sizes.
Hardware: 1B (CPU), 90B (GPU 80GB+)
Mistral Nemo
Mistral's 12B model. Balanced performance, good for conversation.
Hardware: 12B params (~24GB RAM)
Phi-4
Best for CodingMicrosoft's model. Excellent code generation, smaller footprint.
Hardware: 14B params (~28GB RAM)
Qwen 2.5
Best MultilingualAlibaba's model. Excellent multilingual support, large context.
Hardware: 0.5B-72B params
Gemma 2
Google's model. Good reasoning, available in various sizes.
Hardware: 2B-27B params
Quick Reference by Hardware
| Your Hardware | Recommended Model |
|---|---|
| Mac with 8GB RAM | Llama 3.2 1B, Gemma 2 2B |
| Mac with 16GB RAM | Llama 3.2 3B, Mistral Nemo |
| Mac with 32GB RAM | Llama 3.2 8B, Phi-4 |
| GPU 16GB VRAM | Llama 3.2 8B, Qwen 2.5 14B |
| GPU 24GB VRAM | Llama 3.2 70B (Q4) |
| GPU 80GB VRAM | Llama 3.2 90B, Qwen 2.5 72B |
Using Models with OpenClaw
Configure OpenClaw to use Ollama:
# Start Ollama
ollama serve
# In openclaw.json:
{
"models": {
"default": "ollama:llama3.2"
}
}Ready to set up?
Install Ollama and pull a model.