Last Commit writes Día de los Muertos epitaphs for dead GitHub repos with a small Qwen model running in your browser

The page calls itself “An ofrenda for the open source that went quiet.” The button says “Light a candle.” Until you paste something in, the altar reads “Nothing on the altar yet.”

Paste a repository and the card fills in from GitHub’s own records first: whether the project is archived or has simply gone quiet, the span of years it was tended, the people who did most of the tending, and its “last words,” which are the newest human commit message after the merges and version bumps are skipped. For request/request, the “Simplified HTTP request client” that Node projects leaned on from 2011 to 2020, the last words are “README: typo in deprecation message (#3268).” Every card is complete at that point, before any model has loaded.

Then the model writes. Last Commit runs Qwen2.5 Instruct, converted to ONNX by the onnx-community project, through Transformers.js in a Web Worker. On a machine with WebGPU it loads the 1.5B version, about 1.2 GB, with a one-click switch to the 0.5B version at about 480 MB. Without WebGPU it falls back to WebAssembly on the CPU, and only the 0.5B model fits there. Simo tried the bigger one anyway: it “downloaded about 1.8 GB and then ran out of memory inside WebAssembly.” He also tried smaller 8-bit CPU weights, and “they produced garbled text.”

The prompt is four lines and two examples, left-pad and jshint, and the model gets 32 tokens for an epitaph and 96 for a remembrance. In his CPU test, “The first load downloaded 793 MB of weights and took 54.7 s,” and each piece of text took between 18 and 50 seconds after that. The narrated demo on the site is sped up twelve times; the real run took two minutes and fifteen seconds.

The part that makes it more than a novelty is what happens to the model’s answer. According to the repository’s README, “A reply is set aside, and the fact based line stays, if it invents a number that is not in the facts.” The same happens if the reply uses first person, includes a link or an emoji, repeats the prompt, starts chatting, or runs too long, and a panel on the card says why the model lost. Simo’s reason, in his write-up: “I would rather show a plain true line than a confident invented one.”

So Atom gets the model’s 23-token epitaph, written in 29.7 seconds: “It was a hackable text editor, not just a hack, but a true hack, by any standards.” Brackets doesn’t. Its model reply slipped into first person, so the card keeps the line built from GitHub’s numbers: “It gave 7,429 forks a place to begin, then rested in March 2021.” request/request got a 10-token epitaph in 18 seconds, “It was here, waiting for its day,” but its remembrance was set aside for first person, so the card says “Across 2011 to 2020, it was tended by many hands, mikeal first among them.”

Simo is upfront about what the checks can’t catch. They stop invented facts and first person, not dull writing, which is how Atom’s remembrance came out as “Atom was not just a tool; it was a living, breathing ecosystem of features and libraries.” The “gone quiet” label is a rule of thumb too: archived, or no commits on the default branch for a year. “A quiet repository is not always a dead one,” he writes. The code is MIT-licensed, the tests number 48, and he says he built it “with help from an AI coding assistant,” with ElevenLabs used only for the demo voiceover.

Día de los Muertos is November 1 and 2, and the altar starts empty for every visitor. Which repositories people bring to it, and how many of their epitaphs the checks throw out, is the part the repository can’t show yet.