nohamouazine/vertigo-3

Vertigo 3

Small French language model by ParadoxalAI. Own training code, own data, own engine. Runs with Node.js, no GPU, no dependency. Source and bench: https://github.com/nohamouazine/vertigo

Try it in your browser: https://huggingface.co/spaces/nohamouazine/vertigo-3-demo

When it doesn't know, it asks to read: it writes [outil:cherche|…], the engine fetches the intro of the matching French Wikipedia article, and the answer comes with the source.

$ npm run ask "Qui a écrit Les Misérables ?"
Il s'agit d'un roman de Victor Hugo publié en 1862.

Source : Les Misérables · https://fr.wikipedia.org/wiki/Les_Misérables

Run

git clone https://huggingface.co/nohamouazine/vertigo-3 && cd vertigo-3
npm test
npm run ask "Pourquoi le ciel est bleu ?"

Node.js 22.6+. ~60 ms/token on a MacBook, 100 MB of RAM.

Files

• model/ weights (98 MB, int8 + one scale per row), tokenizer, test vectors
• engine/ tokenizer.ts, model.ts, vertigo.ts (generation + Wikipedia tool), ask.ts, test.ts

97M parameters, 12 layers, 1024 context, 16k vocab. Pretrained on French Wikipedia and French web text, then trained to chat and to call its tool. Trained on one PC with an 8 GB GPU.

The weights are our own export format (manifest + int8 shards), read by the engine in this repo. They are not a transformers checkpoint.

Bench

150 French questions, six categories, graded by rules. Vertigo 3: 56.4 / 100. Other small models on the same questions, every answer shown: https://paradoxalai.com/banc

Licence

Apache 2.0, code and weights. contact@weareparis.eu