Can a Local Model Actually Run OpenClaw? Two Honest Tests

I made two videos on running OpenClaw with a local model for zero API cost, one before Qwen 3.5 and one after, and the answer flipped between them. The pitch is simple: no API bill. The catch has always been that the model was not good enough. Here is what changed.

Why people want off the API

Running OpenClaw on a paid model can hit $100 to $200 a month fast, and I have seen people on Reddit blow past $400. My own first setup burned $28 on Anthropic in a few minutes, and swapping to Claude Haiku barely helped. On top of that, Anthropic has been banning users who point OpenClaw at a Claude Code subscription, and Codex only allows it for now. A local model sidesteps all of that. The only question is whether it holds up.

The Qwen 3.5 setup

What changed my mind was a post from Julien, the CTO of Hugging Face, calling Qwen 3.5 a before-and-after moment for local agents: reliable tool calling, stable agentic loops, punching above its weight. So I tested it.

Three commands. Install Ollama with the one-liner from their site, pull the model with ollama pull, and connect it with ollama launch openclaw. No API keys. You pick the 27B or 35B variant based on how much memory you have. Then two tests: writing a simple HTML file passed, and turning it into a working snake game as a React app also passed. The game was genuinely playable. Qwen 3 failed the same tasks a few weeks earlier, so this is a real step up. The model is slower than a paid frontier model, but it runs on your machine for zero dollars against roughly $20 a month on a Codex plan.

The earlier test that failed

Before Qwen 3.5 I ran the same idea live and it did not go smoothly.

That stream connects OpenClaw to Ollama on a 128GB Mac and covers picking a model that fits your memory. Ollama recommended GLM 4.7 Flash. I verified OpenClaw was really using it by asking the model which one it was, and it answered GLM 4.7. Then it failed the coding task: it claimed it saved a file, but the terminal showed the file did not exist. Same setup, weaker model, hallucinated success instead of real work.

What it comes down to

Local removes the API bill, but a local model is slower than a paid frontier model, so most people land on a hybrid. Run local for everyday, non-critical, high-volume tasks on your own hardware, and reach for the cloud when you need speed, parallelism, or a hard task done right.

Watch both videos above to see the difference a better local model makes. To weigh $0 local against your current API spend, openclawdc.com has a calculator for it. I specialize in building hybrid local-plus-cloud setups that cut AI spend. Book a call at cloudyeti.io/meet.