On-Device AI Assistant (Experimental)

4 min read

PinePaper’s AI / Code console has an Assistant that runs a language model entirely in your browser — no account, and nothing you type is sent to a server. Describe what you want, and it builds it on the canvas in one step.

Experimental. On-device models are small, so results vary. For higher-fidelity generation, use the Cloud tab (account) or an external agent via the MCP server.

Where to find it

  1. In the editor toolbar, click </> AI / Code.
  2. The console opens with three tabs: Code · Assistant · Cloud.
  3. Click Assistant (the no-account tab).

If the Assistant tab isn’t visible, your browser doesn’t have a supported on-device engine available — see Requirements.

Using it

  1. Type a prompt, e.g. “a pulsing red star on a dark background.”
  2. Click Generate (or press ⌘/Ctrl + Enter).
  3. The scene is built directly on the canvas — there’s no copy/paste or separate Run step. The panel shows a per-operation summary (e.g. ✓ create star, ✓ execute_generator drawSunburst).
  4. Not happy with it? Press Undo (⌘/Ctrl + Z) to revert and try a different prompt.

Because the model’s output is constrained to PinePaper’s real tools, it can’t produce invalid commands — at worst it makes an odd-but-valid choice (a circle where you wanted a star), never broken code.

Choosing an engine

Use the Engine toggle at the top of the Assistant panel:

Option What it uses Notes
Auto Best available Prefers Browser AI, falls back to PinePaper AI. Default.
Browser AI Chrome’s built-in Gemini Nano Zero download from us — the model ships with Chrome. Fastest to start.
PinePaper AI WebLLM running Qwen2.5-1.5B-Instruct Downloads a ~1.1 GB model on first use (~1.25 GB on disk once cached), then runs offline. Works in any WebGPU browser.

🧠 Models & sources. PinePaper AI runs the open-weight Qwen2.5-1.5B-Instruct (Apache-2.0), served as the WebGPU build Qwen2.5-1.5B-Instruct-q4f16_1-MLC — the ~1.1 GB 4-bit quantization actually downloaded to your browser cache. Browser AI uses Chrome’s built-in Gemini Nano. Open each model card for capabilities, license, and known limitations.

Switching engines re-prepares the model; the first time you pick PinePaper AI, watch the progress bar while the ~1.1 GB model downloads. After that it’s cached on your device and runs offline.

Languages

If your editor language isn’t English, your prompt is translated to English on-device (via the browser’s built-in translator, when available) before the model runs — the small models understand English best. The generated scene is the same regardless of prompt language.

Privacy

Everything happens on your device:

  • Your prompt is not sent to any server.
  • The model runs locally (Chrome’s built-in model, or one cached in your browser).
  • No account is required.

Requirements

  • Browser AI (Gemini Nano): Chrome on desktop (Windows 10/11, macOS 13+, Linux) with built-in AI enabled. If the Assistant says no provider is available, enable “Prompt API for Gemini Nano” at chrome://flags, relaunch, and confirm the model is downloaded at chrome://on-device-internals.
  • PinePaper AI (WebLLM): any browser with WebGPU (Chrome/Edge 113+, Safari 18+) and ~1.25 GB of free browser storage for the model cache (a one-time download).
  • Translation needs Chrome’s built-in Translator/Language Detector (desktop Chrome).