Running cloud models like Claude or GPT inside OpenClaw means token fees are an ongoing expense. Google has just published an official tutorial showing how to hook the local Gemma 4 model into OpenClaw in three steps and run it at zero token cost.

This tutorial lays the process out clearly, and also covers what it is good for and where it falls short.

What This Tutorial Helps You Do

Use Gemma 4 as OpenClaw's backend model, running on your own machine, without spending a cent on tokens. It suits simple scenarios like briefing generation, meeting transcription, and scheduled tasks.

Before You Start

Hardware Requirements

What You Need

• Ollama (local model runtime)
• OpenClaw (the Agent framework)
• About 10 minutes of setup time
• Zero cost

Step-by-Step Instructions

Step 1: Install Ollama

Go to https://ollama.com/download and download the installer for your platform.

macOS users can also use Homebrew:

brew install ollama

Once the installation is done, start Ollama:

ollama serve

Tip: Make sure the Ollama service keeps running in the background; the later steps depend on it.

Step 2: Download the Gemma 4 Model

The official recommendation is the 26B A4B version (MoE architecture, fewer parameters actually activated, faster):

ollama pull gemma4:26b-a4b

If your hardware falls short, you can pick a smaller version:

ollama pull gemma4:12b-a4b
ollama pull gemma4:4b

Tip: The 26B A4B version download needs roughly 15-20GB of disk space; make sure you have enough storage.

Step 3: Launch OpenClaw via Ollama

This single command installs OpenClaw automatically and starts it with Gemma 4 as the backend:

ollama run gemma4:26b-a4b

Then, in OpenClaw's configuration file, point the model backend at the local Ollama service. The exact configuration depends on your OpenClaw version; usually you pick 'Ollama' as the provider in the model settings and enter http://localhost:11434 as the endpoint.

Verifying the Result

Once it is up, send a simple message in OpenClaw to test. If you get a reply, the local model is successfully connected.

You can test a few scenarios to gauge the quality:

• Briefing generation: give it a block of text to summarize
• Meeting transcription: paste in meeting notes and have it organize them
• Scheduled tasks: set up a periodic reminder or data rollup

Where It Fits and Where It Does Not

Good Fits

• Text processing tasks like briefs, summaries, and meeting minutes
• Background automation tasks on a schedule
• Everyday workflows that do not need complex reasoning
• The exploratory phase, when budget is tight and you just want to get the pipeline working

Poor Fits

• Complex programming tasks (the gap to top models like Opus is obvious)
• Tasks that require long-context understanding
• Production environments with strict security requirements

Note: OpenClaw founder Peter Steinberger has publicly advised against cheap small models, because small models are more susceptible to prompt injection attacks. Evaluate the risk yourself when handling sensitive data.

Cost Comparison

Some users have done the math: if your daily workload is just briefings and transcription, a Mac Studio pays for itself in token savings within 3 months.

FAQ

• Q: Can a Mac Studio M4 Pro 48GB run the 26B version? A: Yes. Actual VRAM usage is about 16GB, so the machine still has headroom.

• Q: How is the response speed? A: Simple questions run smoothly; it slows down when the context gets long or deep thinking is enabled. On an M4 Pro, everyday use is acceptable.

• Q: How much worse is it than Claude? A: The gap in reasoning ability is obvious, especially in tool calling and long contexts. But it is good enough for simple tasks.

• Q: Is it safe? A: Small models are weaker at resisting prompt injection. If you handle sensitive data, use a stronger model.

References

• Ollama download: https://ollama.com/download
• Gemma 4 official announcement: https://ollama.com/library/gemma4
• OpenClaw official documentation: https://openclaw.dev