Gemma 4 inside Obsidian — no server, no API key, no separate process

There's an Obsidian plugin that runs Gemma 4 E4B inside the app using LiteRT-LM and WebGPU, with nothing else running alongside it. Keeping Ollama running next to Obsidian means the AI feature crashes silently every few days; you find out only when it stops responding. Embedding the model inside Obsidian's own process removes that layer. Open the app, the model is there.

• Obsidian runs on Electron, which is Chromium packaged as a desktop app. LiteRT-LM (Google's rebranded TensorFlow Lite) is built for this kind of constrained browser environment: phones, embedded hardware, and the WebGPU runtime that Electron exposes.
• No API key, no server, no second process to keep alive. The failure surface disappears because there's nothing to fail.
• The model runs offline, no subscription required. For anyone keeping sensitive notes in Obsidian, that changes the threat model.
• My open question is context depth. 4B parameters run fast, but context depth on consumer WebGPU hardware is a different question.

Implementation details in the wiki and on GitHub.

Written under the word limit,
Zhenia