OpenClaw + Ollama: Run AI Models Locally (Hardware Guide)

Local models are why most people pick OpenClaw over a cloud-only assistant. The tradeoff is hardware — you bring the compute, you get the privacy and zero per-token cost.

This page is the practical guide: what hardware runs what, how to set it up with Ollama, and when local actually beats cloud.

TL;DR

If you're under 16GB, you're better off paying $5/month for cloud API access than fighting OOM errors. If you're at 24GB unified memory or above, local is competitive.

Quick setup with Ollama

1. Install Ollama

# macOS
brew install ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

Start it:

ollama serve

2. Pull a model

ollama pull llama3.1:8b          # 4.7 GB, good general-purpose
ollama pull qwen2.5:14b          # 9 GB, better tool-calling
ollama pull qwen2.5-coder:7b     # 4.7 GB, for code tasks

Browse the full library at ollama.com/library.

3. Configure OpenClaw

Edit ~/.openclaw/openclaw.json:

{
  "providers": {
    "ollama": {
      "baseUrl": "http://localhost:11434/v1",
      "models": ["llama3.1:8b", "qwen2.5:14b"]
    }
  },
  "agents": {
    "default": {
      "model": "ollama/llama3.1:8b"
    }
  }
}

Restart and verify:

openclaw gateway restart
openclaw agent --message "Test — are you running locally?"

If the response comes back, you're running fully local.

Hardware tiers in detail

Tier 1 — Light local (8 GB unified / VRAM)

Realistic models: Llama 3.1 8B, Mistral 7B, Qwen 2.5 7B.

These handle conversation, summarization, and simple tool calls. They struggle with multi-step reasoning, complex coding, and structured output. Token rate on Apple Silicon M2 8GB: ~20-30 tok/sec.

Good for: a personal Telegram bot, daily journal summarization, basic Q&A.

Tier 2 — Useful local (16 GB)

Realistic models: Qwen 2.5 14B, Llama 2 13B, Qwen2.5-Coder 14B.

The minimum where local feels "good enough" for daily use. Tool calling is reliable, code generation is competent, multi-step reasoning works on focused tasks. ~10-20 tok/sec on M-series.

Good for: a serious coding assistant, multi-skill agents, MCP-driven workflows.

Tier 3 — Strong local (24 GB unified memory)

Realistic models: DeepSeek-R1 32B (4-bit), Llama 3.3 70B (heavy quantization).

The official OpenClaw docs describe 24 GB as "suitable only for lighter prompts with higher latency" — translation: you can run good models, but expect 5-10 tok/sec and longer time-to-first-token on big contexts.

Good for: a primary daily-driver agent that replaces cloud usage for most tasks.

Tier 4 — Production local (32-64 GB+)

Realistic models: DeepSeek-R1 32B full, Llama 3.3 70B, Mixtral 8x22B.

Cloud-competitive quality for most tasks. Multi-agent setups become viable here — one model running, multiple agents sharing it. 5-15 tok/sec.

Good for: 24/7 multi-agent orchestration, fully air-gapped deployments.

When 24 GB isn't enough

The official docs note: "≥2 maxed-out Mac Studios or equivalent GPU rig (~$30k+)" is the recommended hardware for serious multi-agent local stacks. If your goal is replacing a team's worth of Claude/GPT-4 usage with local infrastructure, that's the realistic price floor.

Mixing local and cloud

Most OpenClaw users end up with a hybrid:

{
  "providers": {
    "ollama": { "baseUrl": "http://localhost:11434/v1" },
    "anthropic": { "apiKey": "${ANTHROPIC_API_KEY}" }
  },
  "agents": {
    "personal": { "model": "ollama/qwen2.5:14b" },
    "research": { "model": "anthropic/claude-sonnet-4-5" }
  }
}

The personal agent handles your private data on-device. The research agent calls Claude when you need deep reasoning. Same gateway, same channels, both available.

For the cloud-orchestrator + local-text-workers pattern (a cloud model orchestrates, local models do bulk text work), see the OpenClaw local models guide.

Embeddings (memory search)

Local embeddings are separate from chat models. OpenClaw's memory search needs a dedicated embedding model:

ollama pull nomic-embed-text     # Fast, 274 MB, good default

{
  "providers": {
    "ollama": {
      "embeddingModel": "nomic-embed-text"
    }
  }
}

If you're on Apple Silicon, native MLX embedding support via oMLX is tracked in the OpenClaw roadmap.

Common gotchas

• Ollama not bound to network: by default Ollama listens on localhost only. If OpenClaw runs in Docker, set OLLAMA_HOST=0.0.0.0 and use the host IP in baseUrl.
• Proxy interference: as of v2026.5.19 there's an open bug where SSRF defenses ignore NO_PROXY when calling local Ollama embeddings. Workaround: disable the proxy for embedding traffic.
• Cron model preflight: if your cron job's primary model is a local Ollama target and the local server is offline at preflight time, the entire run is skipped (cloud fallbacks ignored). Tracked bug — for production cron, prefer cloud as primary with local as fallback.
• Reasoning models (<think>/<final> tags): some local reasoning models leak reasoning content. Configure your provider to strip these tags, or use a model without explicit reasoning channels.

→ Ollama model library · System requirements · OpenClaw + Ollama setup · Local models guide