"Why are you making such a small model?" is the most common question we get. There are four good answers and none of them are about beating GPT-5.
1. It runs anywhere
Tilelli is 39 MB. It fits on a Raspberry Pi. It loads in under a second on a five-year-old laptop. The
inference loop is single-threaded NumPy if you want it to be. There is no API key, no cloud quota, no rate limit.
A model that requires a remote GPU to answer a question is a model that can be paywalled, throttled, deprecated
overnight, or quietly retrained on your conversations. A model that fits in 39 MB and runs on the CPU you already
own can't be any of those things.
2. You can read all of it
The entire weight tensor of Tilelli is small enough to inspect by eye. The architecture is two hundred lines of PyTorch. The training recipe is one shell script. You can read everything that determines its behavior in an afternoon.
That's not nostalgia for "simple software." It's the precondition for auditing the model: knowing what it does, what it can't do, what it was trained on, and what it would refuse.
3. You can re-derive it cheaply
Reproducing GPT-4 is impossible — it would cost tens of millions of dollars and you'd need to source the data. Reproducing Tilelli v0.1 costs less than a movie ticket. A graduate student, an independent researcher, or a curious teenager can train it from scratch on one rented GPU and verify every number on this site.
4. It's a clean substrate for research
Small models make experiments fast. Every metacognition experiment in our postmortem ran in minutes, not days. We could try, fail, splice, control, and disprove in a single afternoon. That speed is what made it possible to publish a negative result instead of waving hands at a confound.
What we don't claim
Tilelli will not write your dissertation. It will not refactor a microservice. It will not pass the bar. It is a small research model that does interesting things at a price you can afford. If you need a frontier model, use a frontier model. If you need a model you can own, Tilelli is the kind of thing we should be making more of.