Is this plugin open source? Yes — MIT.
Is this plugin completely free? Yes — no paid tier, no account, no telemetry.
Is this plugin vibe-coded beyond the author’s ability to comprehend how it works? No.
Community Directory:https://community.obsidian.md/plugins/gemma-litert-wiki
Every local-AI plugin I tried needed a second program: install Ollama or LM Studio, start it, keep it running, point the plugin at localhost:11434. I wanted to know whether that second program was actually necessary, or just the way everyone had done it.
It isn’t necessary. Gemma 4 E4B LLM Wiki loads Gemma 4 E4B into Obsidian’s own renderer process through LiteRT-LM and WebGPU. There is no server, no API key, no account, and nothing to keep alive. You install the plugin, download the model once, and that is the whole setup.
What it does with the model
It builds Andrej Karpathy’s LLM-wiki pattern (https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f) out of your notes — a gemma-wiki/ folder holding one card per note (summary, tags, key points, salient mentions, a self-rated confidence), an index.md catalog, concept pages built across every card sharing a tag, and an append-only activity log. All of it plain Markdown you can read, edit, grep and version. No database.
Then you can talk to it. The chat panel grounds either in the note you have open or in the whole wiki through index-first retrieval, and the plugin lists the sources itself rather than asking the model to cite them. Built-in 
skills turn a note into quizzes, flashcards, or a list of gaps; drop a markdown file in gemma-wiki/skills/ and you have your own.
Two things I want to be upfront about
Nothing is written without your approval. Every page, every rewrite, every tag suggestion is rendered in full in a preview first. The two commands that touch a raw note you wrote — Improve formatting, and Suggest tags & links — always preview.
Everything retrievable derives from a note you wrote. Saved answers and saved conversations live in the wiki folder but are never indexed, retrieved, linted or scanned. An answer is output, never input — a chat sitting in the index would be indistinguishable from a page.
What it costs you
A one-time ~3 GB model download, and ~20–31 MB for the WebAssembly runtime. After that it never touches the network again. Desktop only: it needs WebGPU and the disk space, so there is no mobile version and structurally cannot be one.
It is one model — Gemma 4 E4B — and you cannot bring your own. That is the trade for having nothing to install alongside it.
See it before installing anything
The full walkthrough runs in a browser, no download: https://gemma-wiki-demo.vercel.app
Source, benchmarks and the honest comparison table (including the rows where this approach is worse): https://github.com/itsyuimorii/obsidian-gemma4-litert-wiki
Happy to answer anything here — especially if you try it on a vault shaped differently from mine.