Can a Four-Bit Local Model Actually Ship a Feature? Pi Says Yes

I made a video testing Pi, the minimal coding agent people have been raving about, with a local model. A four-bit model running on a laptop sounds like it should struggle to edit a real file, but in my test it added a working dark-mode toggle without help. Here is the gist, the setup, and why the agent is the part that mattered.

Why Pi over a bigger agent

Pi is a minimal terminal coding agent with just four built-in tools. No MCP support, no sub-agents, no permission loops, no plan mode, no built-in todos, no background bash. On a cloud model that trimming barely matters, but on a local model it changes the outcome, because a smaller model has less room to absorb a heavy system prompt and long tool chains.

So Pi keeps the surface small and lets the model spend its limited attention on the actual code. The people pushing it lately include the creator of Redis and the CTO of Hugging Face, mostly for running local models offline.

The setup I used

I served the Qwen 3.6 27B four-bit model through llama.cpp on my Mac, on port 8081, and pointed Pi at it. llama.cpp is the local runtime that hosts the model, and Pi connects to it as a provider. If you already run models with Ollama or MLX-LM, Pi can use those instead, so you are not forced to change your stack.

Starting Qwen 3.6 27B locally and selecting it inside Pi is a short sequence, and the video walks through each step.

The result

The task was a working dark-mode toggle in a real HTML file, run on a four-bit Qwen. Four-bit means the model is quantised down for a smaller memory footprint, which is exactly the kind of model you run when you want local coding to fit on a normal machine. Even at four bits this is a dense 27B, and in the run the inference felt fast, not sluggish.

Pi wrote a plan, used CSS custom properties for theming and JavaScript to toggle the class, and saved the file. The toggle worked. That is the signal I care about: a compressed local model, a minimal agent, and a feature that actually works at the end.

Watch the full install and demo in the video above. The models.json config is linked in the description. For more local-AI setup guides, openclawdc.com has walkthroughs on running models on your own machine.

I specialize in helping teams wire up local coding agents that ship real features. Book a call at cloudyeti.io/meet.