@aparte/titler

The runtime for aparte-titler models: a title of 3 to 6 words for a conversation, copied verbatim from its first message, in the browser or in Node, with no API call. One file, no dependency, ~6 KB minified, 6–7 ms per title on one CPU core.

This package is the runtime only. It reads the .bin model files published at huggingface.co/apartejs/aparte-titler. If you just want it to work, install @aparte/titler-latin instead: same runtime, default model bundled (17 languages, 133 KB).

Install

npm install @aparte/titler

Use

import { Titler } from "@aparte/titler";

// any model from https://huggingface.co/apartejs/aparte-titler
const url = "https://huggingface.co/apartejs/aparte-titler/resolve/main/titler-v1.1-latin-int3.bin";
const titler = new Titler(await fetch(url).then((r) => r.arrayBuffer()));

titler.title("Write me a cover letter for a junior data analyst position at a bank");
// -> "cover letter junior analyst position bank"

titler.title("Salut ! Tu peux me donner une recette de pain sans gluten facile pour ce week-end ?", 4);
// -> "pain gluten facile week-end"   (a budget of 4 words instead of 6)

titler.words("Can you explain how photosynthesis works in plants?");
// -> { words: [{ word: "Can", start: 0, end: 3, score: 0.11 }, …, { word: "photosynthesis", start: 19, end: 34, score: 1.00 }, …], … }

In Node, read the file instead of fetching it:

import { readFile } from "node:fs/promises";
const buf = await readFile("titler-v1.1-fr-int3.bin");
const titler = new Titler(buf.buffer.slice(buf.byteOffset, buf.byteOffset + buf.byteLength));

Give the model only the user's first message — never a system prompt or the conversation history: it picks the most salient words of whatever it is given.

Which model file?

| file | covers | int3 size | |---|---|---| | titler-v1.1-latin-int3.bin | 17 European languages (en, fr, es, de, pt, it, nl, pl, sv, da, fi, cs, ro, no, hu, hr, lt) | 133 KB | | titler-v1.1-latin-mini-int3.bin | the same 17, smaller vocabulary, −0.5 point on average | 96 KB | | titler-v1.1-efigsp-int3.bin | en, fr, es, de, pt, it | 77 KB | | titler-v1-<lang>-int3.bin | one language | 40 KB |

Every model also exists in fp32, int8 and int4. int3 is the recommended precision: same score as fp32 on every benchmark. Scores, sizes and charts: huggingface.co/apartejs/aparte-titler.

What it does, exactly

A 2-layer transformer encoder (58k parameters for one language, 141k for seventeen) scores every word of the message; the title is the best-scored words, kept in message order. The words are always copied from the message — it never rewrites, never invents, never corrects a typo. It also cannot execute anything: the worst an injected instruction can do is a poor title.

Format and versions

The runtime reads model files of format: 1. A future format 2 (new architecture) will ship with a new major version of this package that still reads format 1 files. Models and runtime follow the same versioning: 1.x keeps the file format, 2.0 changes it.

License

MIT © 2026 Paul Richez. The models have their own cards and licenses on Hugging Face.