This is a submission for the Gemma 4 Challenge: Build with Gemma 4

What I Built

Sowser is a spatial canvas browser for Windows — instead of a strip of tabs, every website lives as a draggable, resizable live card on an infinite canvas. You pan, zoom, group, and connect pages the way your brain actually thinks, not the way a tab bar forces you to.

The problem is simple: traditional browsers trap you in a single column of tabs. When you are researching, comparing, or debugging across many sources, you constantly alt-tab and lose context. Sowser solves this by giving every page space — you see everything at once, arrange it meaningfully, and save that layout as a workspace to return to later.

Built with WPF and Microsoft Edge WebView2 on .NET 8. Cards are real live pages, not screenshots. The canvas supports sticky notes, image clips, connection lines between cards, a command palette (Ctrl+K), minimap, tracker blocking, per-card browser profiles, and full workspace save and load.

Gemma 4 powers the centrepiece feature: AI Smart Organize.

When you click AI Smart Organize, Sowser:

1. Collects every open card's title and URL
2. Sends them to Gemma 4 running locally via Ollama
3. Receives a JSON grouping of cards into named semantic clusters
4. Repositions cards into colour-coded columns on the canvas automatically

The canvas goes from chaos to clarity in one click — privately, locally, with no data leaving your machine.

Demo

🔗 github.com/noisyboy08/TREE-TABS

Watch the demo video in the repository (demo.mp4) to see:

• Multiple browser cards open across different topics
• One click on AI Smart Organize
• Gemma 4 grouping them into labelled colour-coded columns instantly
• Zooming out to see the full spatial layout

Code

🔗 github.com/noisyboy08/TREE-TABS

Core files for the Gemma 4 integration:
Models/GemmaSettings.cs — model config, endpoint, local vs cloud toggle
Services/GemmaService.cs — HTTP client to Ollama or Gemini API fallback
Services/GemmaOrganizeService.cs — parses groups, repositions cards on canvas
MainWindow.FeaturePack.cs — wires the AI Smart Organize menu handler

To run the AI feature locally with no API key:

ollama pull gemma3:4b
ollama serve

Then open Sowser, load some tabs, and click AI Smart Organize.

How I Used Gemma 4

Model chosen: Gemma 4 E4B — 4 billion effective parameters, running locally via Ollama.

Why E4B specifically

The Smart Organize feature needs to run fast, silently, and without sending your browsing history to any cloud. E4B is the exact right fit:

• Fast enough to feel instant on a mid-range laptop. First call loads the model into memory. Every call after that returns in one to three seconds.
• Smart enough to semantically cluster URLs and page titles into meaningful groups. This is not keyword matching — it is genuine topic understanding.
• Local by default — your tabs never leave your machine. This matters deeply for a privacy-minded browser tool.
• No GPU required — making it accessible to the widest possible Windows user base.

I tested the 2B model and found it produced noisier groupings that sometimes merged unrelated topics. The 31B Dense model is overkill for a clustering task and would feel sluggish in an interactive desktop UI. E4B hits the right balance between speed, accuracy, and hardware accessibility.

What Gemma 4 actually does inside Sowser

The prompt sent to the model is strict and unambiguous:

• Receive a JSON array of open tab titles and URLs
• Return only a JSON array of groups
• Each group has a name, a hex colour, and the list of URLs belonging to it
• No explanation, no markdown, no preamble — just structured data

Gemma 4's instruction-following quality is what makes this reliable in a production desktop app. The app strips any accidental markdown fences, parses the JSON, then immediately repositions and recolours cards on the canvas. If the model output is malformed for any reason, the app catches the error, shows a toast, and never crashes.

Why this only works because of Gemma 4

A browser that understands what you are looking at and organises it for you — locally, privately, instantly — is only possible because models like Gemma 4 exist at this size and capability level. That is the real unlock. Not AI as a cloud API you pay per token. AI embedded directly into a native desktop tool, running on your own hardware, with zero data leaving the device.

Sowser was already a spatial browser. Gemma 4 made it a spatial browser that thinks.

Think in space, not in tabs.