Tiny Models
Who needs a trillion parameters? Here the models fit in your palm. All real ONNX, all in your browser, all SOTA (on our own benchmark).
- Run ▸
CoinFlip‑1P
1 param
The undisputed champion of yes/no questions.
- Accuracy
- 50.0% (SOTA)
- Hallucinations
- 0
- Size
- 4 bytes
- Run ▸
DiceRoll‑6P
6 params
6× the parameters, 3× worse accuracy. Scaling laws, baby.
- Accuracy
- 16.7% (SOTA)
- Hallucinations
- 0
- Size
- 24 bytes
- 🤗 Hugging Face Run ▸
Tavla‑1.6K
1601 params
1,601 parameters that taught themselves backgammon by self‑play. Your move.
- vs random
- 99.7%
- vs heuristic bot
- 75.3%
- Size
- 6.6 KB
- Run ▸
Okey‑257
257 params
257 parameters, 3 bots, one rack. Classic Turkish okey, learned from bot self‑play. Indicator’s up.
- vs 3 random bots
- 93.4%
- vs 3 greedy bots
- 24%
- Size
- 1.4 KB
- Run ▸
Kedi‑27K
27044 params
Hide the toy under a cup and shuffle. The cat guesses from nothing but the pixels its camera sees. How fast can you shuffle? The real test table is on the page.
- Accuracy (held‑out)
- 100%
- Random guess
- 33.3%
- Size
- 111 KB
- Coming soon
RockPaperScissors‑3P
3 params
Trained on three hands. Ties expected.
- Accuracy
- 33.3% (expected)
- Training
- in progress
- Coming soon
Magic8Ball‑20P
20 params
20 answers, 0 understanding. A true large language model.
- Answers
- 20
- Understanding
- 0