Your models. Your hardware.

Free AI, without the mystery.

Connect open-weight coding models to build websites without per-request AI charges on your own infrastructure.

Back to builder

Model connection

Not connected

Choose a coding model

Smaller models fit more machines. Stronger models need more memory.

Practical default

Qwen 2.5 Coder · 7B

A smaller dedicated coding model for everyday websites. A sensible starting point when hardware is unknown.

4.7 GB weights · additional memory needed

Official model details ↗

Stronger agentic option

Qwen3 Coder · 30B

Better suited to complex multi-file tasks on a capable workstation. 30B total parameters; about 3.3B active.

19 GB weights · context adds memory

Official model details ↗

High-memory option

Qwen3 Coder Next

Agentic coding model for substantial local hardware. Not an appropriate default for phones or small laptops.

52 GB quantized weights · context adds memory

Official model details ↗

Lightweight fallback

Qwen 2.5 Coder · 3B

For smaller machines and simple edits. Complex reasoning and large projects may exceed this model’s capabilities.

1.9 GB weights · additional memory needed

Official model details ↗

Set up your local model

1. Install Ollama on your computer

The website hosts the interface, not the model weights. Run inference on your own computer or a machine you control.

Get Ollama

2. Allow this app and pull your model

launchctl setenv OLLAMA_ORIGINS "https://ai-website-studio.genmb.com"
# Quit and reopen Ollama, then:
ollama pull qwen2.5-coder:7b

Restart Ollama after setting allowed origins. Do not allow every website or expose an unprotected Ollama port to the internet.

3. Test, then build

Use Test real inference. It discovers installed models and sends a real prompt. A status badge is never a substitute for a successful response.

Actual limits, plainly stated

Local inference has no per-request API charge, but uses memory, compute, disk space, electricity, and download bandwidth.
There is no artificial chat quota here. Context length, output size, and your hardware constrain a request.
Built-in storage limits each value to 100 KB. Larger private projects, drafts, and checkpoints are split into smaller records. Shared limits of 600 reads and 180 writes per minute still apply, so large edits can take more requests. Public snapshots must fit in one record.
React previews use Sandpack’s external bundler. Static previews run in your browser. Neither runs arbitrary backend servers.
Phones and tablets need a reachable, secured inference server on another machine. “localhost” on your phone is not your laptop.
Compatible endpoints must permit browser requests. Provider authentication can be handled in your own trusted proxy; do not place secret API keys in project files.

Connection troubleshooting

The model is not installed

Run the pull command, then discover models. A smaller fallback is offered only when it is already installed.

The browser says Failed to fetch

Verify Ollama is running. Configure OLLAMA_ORIGINS for this app’s displayed origin and restart it. Allow local-network requests in your browser. If HTTP is blocked, use an authenticated HTTPS proxy or run the builder locally.

Generation is too slow or runs out of memory

Choose a smaller model, shorten the request, reduce large files, and close memory-heavy applications. Model-file sizes are not total runtime-memory requirements.

I need remote or paid-provider inference

Use an OpenAI-compatible endpoint with CORS enabled or a trusted, authenticated proxy. Basic local inference does not require a paid key. Remote free tiers may impose their own limits.