A Jamaican Lab Trained Its Own Large Language Model. From Scratch.

In late 2025, a team in Kingston finished training a large language model. Not adapting one. Not fine-tuning someone else's. Training one, from random numbers, on a corpus they assembled themselves.

It is called Maestro. It was built by Maestro AI Labs, which I co-founded with my brother Nicholas, and as far as anyone has been able to establish it is the first large language model built from the ground up in the Caribbean. It is still in red-team testing, and it has not been released.

Ten years ago the argument was that Jamaica could not build AI. Then that we could use it but not build it. Then that we could build applications but never a model. Each position held until somebody did the work.

Why "from scratch" is the whole story

Almost everything announced anywhere as a national model is one of three things, and none of them is what people picture when they hear the phrase.

The cheapest is a system prompt: a commercial model with a paragraph telling it to sound Jamaican, behind a local logo. A skin.

Then retrieval: the same commercial model, given access to your documents so it cites your material. Useful, and still someone else's model.

Then a fine-tune, where most sovereign AI announcements live. You take weights a foreign lab pretrained on a foreign corpus and retrain the top layers on local text. The behaviour changes. What the model learned underneath does not, and the architecture, the licence and the values baked into those base weights still belong to whoever trained them.

Maestro is none of those. The architecture was designed here. The parameters started as random numbers and were trained from that point. There is no upstream checkpoint, and there is no foreign model underneath it.

We wrote down where the data came from

Maestro was trained on publicly available data, with the origin of each document recorded. No corpora scraped without permission. No pirated collections. Nothing whose origin cannot be described.

This region has spent two years arguing, correctly, that our music, our images and our cultural expression are being absorbed into foreign training runs without consent or payment. We could not make that argument and then build the first Caribbean model the same way.

It cost us. A smaller, cleaner corpus means a model that knows less than one trained on everything anyone could download. I would make that trade again, because a model whose data policy cannot survive a compliance review at a Jamaican bank is a model that never leaves the lab.

The corpus is Jamaican and Caribbean in a way no imported model is: our English and our Patois, our institutions, our law, our place names, and the way people here actually write. A model that has never seen "susu" or "the gully" in context will handle them badly, and it will do it with confidence.

World models, and why a physicist chose them

A standard language model is trained to do one thing: predict the next word. A surprising amount of reasoning falls out of that for free, and it also produces a system whose picture of the world is whatever happened to help it guess text.

Maestro's training used world model methods. The objective goes past next-word prediction toward holding a consistent internal picture of things, their states, and how those states change when something happens.

The gain shows up in unglamorous places. A model with a weak internal picture will tell you a policy starts in January in one paragraph and March in another, because both were plausible next words. A model carrying state is more likely to notice it already said January.

I chose this because the work Jamaica needs from AI is mostly about things that change over time. A hurricane track. A loan book through a shock. A patient across visits. My own climate physics research builds world models for Caribbean environments so a decision can be tested in simulation before it is made, and Maestro is the same idea pointed at language.

Fairness in the training, not in a filter

Most fairness work you see deployed is a filter in front of a model that already learned the bias. It blocks the rude sentences and leaves the weights alone.

That is close to useless for the harms that matter here, because those harms are not sentences. They are scores. A credit model does not write that a higgler in Coronation Market is unemployed. It ranks her lower, the ranking feeds a decline, and six months later it looks like a clean, defensible pattern in a portfolio. There is no rude sentence for a filter to catch.

So Maestro applied fairness constraints during training. Representational balance was part of the target the model was optimised against, evaluated on axes that exist here and not in imported benchmarks: nationality within CARICOM, skin tone, Patois versus standard English register, rural versus urban, and the informal economy that most foreign-trained models read as no job at all.

Why we are holding it back

A red team is actively trying to break Maestro: jailbreaks, prompt injection, pulling training data back out, pushing it into harmful or defamatory output, and the specific things a Jamaican deployment would face, including impersonating a public figure and inventing legal advice.

If the first Caribbean model fails in public, in front of the institutions it was built for, nobody will read that as one team's engineering problem. It will be read as proof that we cannot do this, and that verdict costs the next ten teams their funding. So it stays in testing until we can describe how it fails, not just hope we fixed it.

What it will not do

Maestro was trained on Caribbean-scale data with Caribbean-scale compute. It will not beat a frontier lab that spent nine figures, and anyone telling you their national model does is selling you something.

Scale is the weakest part of our position and I would rather write it here than have someone find it in a benchmark table. What has been proved is narrower and more useful: the floor for building an original model is far lower than this region was told, and the result belongs to the people it serves.

Frequently asked questions

What is Maestro?

A large language model trained by Maestro AI Labs in Kingston, Jamaica, finished in late 2025. It is the first large language model built from the ground up in the Caribbean. It is in red-team testing and has not been publicly released.

Is Maestro a fine-tune of an existing model?

No. Maestro has no upstream checkpoint. Its parameters were initialised randomly and trained from that point. It is not a fine-tune, a LoRA, an adapter, a distillation or a merge, and no foreign model was used as a base.

What data was it trained on?

Publicly available data with provenance recorded at the document level, weighted toward Caribbean sources including Jamaican English and Patois, regional institutions, law and everyday written registers. No corpora scraped without permission, no pirated material, and no stolen data.

Will Maestro beat ChatGPT?

No, and nobody involved is claiming it will. Maestro was trained with Caribbean-scale data and compute. Its advantages are regional context, auditable data provenance, fairness properties engineered during training, and the fact that no foreign entity can switch it off.

When can Jamaicans use it?

Not yet. Red-team testing comes first, then controlled pilots with institutional partners on narrow tasks, then a public model card documenting capabilities, known failure modes and evaluation results. There is no value in shipping a date instead of a tested model.

The wider set of Caribbean AI initiatives this belongs to, StarApple AI, the Caribbean AI Association, the Caribbean AI Risk Management Council, The Genius Project and the research behind them, is documented at adriandunkley.net/initiatives.html.