Clue Vibes: Taboo, except the tiny model is the one guessing
That flip is the whole game, as you already know the answer: the challenge is figuring out how to talk to a tiny model under weird constraints so it lands the word from one short sentence.
The setup
The game pulls a random card from a deck of 7,237 generated Taboo cards. Each card has a target word and five banned clue words. After filtering out targets that are too short or not simple enough, 7,101 cards stayed in play. A round gives you the target, the five banned words, a 30-second timer, and a 280-character box (basically the length of a tweet!).
You write a clue without using the target or any of the banned words. When you submit, Tiny Aya, a 3.35B multilingual model by Cohere Labs, gets only your clue and tries to name the target. Guess wrong and you can ask it to try again; the retry prompt tells it which words it already missed. You get up to three model guesses per round. If the timer runs out, the app sends whatever you've written, as long as it isn't empty or invalid.
For "airport" with plane, flight, terminal, luggage, passport all banned, you have to find another road into the concept - that's where the fun is!
Why this is actually a semantics exercise
The game quietly trains a skill. To get a small model to guess a word from a single clue, you have to be genuinely precise about its meaning. You learn what the model's concept space looks like, which descriptions are specific enough, and which clever clues are too oblique for something this size. A clue that works on a friend can sail right past a tiny model, and a clue you thought was too plain can land instantly. You start writing for how this particular model carves up meaning, and that's a real skill dressed up as a party game.
The smallness of the model isn't hidden or apologized for: it is the source of the comedy and the design problem at the same time.
How it runs
The model is onnx-community/tiny-aya-global-ONNX, adapted from CohereLabs/tiny-aya-global, loaded in the browser with Transformers.js and run on WebGPU. The Space is CPU Basic, because it does no inference at all. It serves the app, picks cards, validates clues, and scores guesses. The first round downloads roughly 2.3 GB of model files into your browser cache, which the How To Play modal warns you about up front.
A couple of honest limits: validation is simple string and variant matching rather than full lemmatization, the cards are English only for now, and Chrome or Edge give the best WebGPU experience. The banned words were all generated offline, so gameplay stays fast and predictable.
Hackathon fit
This is a Thousand Token Wood entry, and the AI is load-bearing in the clearest possible way: the model itself is the player! Take it out and there's nothing to play. It's aiming at a few badges: off-the-grid, since there's no cloud inference and the model runs locally; custom UI, since the frontend is hand-built HTML, CSS, and JS through Gradio Server; and Tiny Titan, since Tiny Aya sits well under the 4B line. Codex was an active collaborator on the backend, the frontend, the Transformers.js wiring, and the deployment.
The design problem that took the most work wasn't the model or the backend, but the game feel: the timer, the retries, the validation, and the feedback that makes a wrong guess funny instead of frustrating. Once a wrong first guess became something you wanted to recover from rather than a dead end, the game clicked.
Play it here: https://huggingface.co/spaces/build-small-hackathon/clue-vibes
Michele Ciletti Pre-doctoral Researcher in Computational Linguistics, University of Foggia, Italy
