Maestro: We Built the Caribbean's First Large Language Model From Scratch

Every other model in this region is somebody else's model wearing local clothes. Maestro is not. We trained it from nothing, on data we can account for, with fairness built into the objective instead of stapled to the output.

In late 2025, the team at Maestro AI Labs finished training Maestro. It is a large language model, and it is the first one built from the ground up in the Caribbean.

That sentence is doing more work than it looks like it is doing, so let me be precise about what it claims and what it does not.

What "from scratch" actually means

There is a spectrum of things people call "our national model", and almost all of them sit at the cheap end of it.

At the cheapest end is a system prompt. You take a commercial model, put a paragraph in front of it telling it to sound Jamaican, and put your logo on the interface. This is a skin.

One step up is retrieval. You keep the foreign model and give it access to your documents, so its answers cite your material. Genuinely useful, and still someone else's model.

Next is a fine-tune, and this is where most "sovereign AI" announcements live. You take an open-weight model that someone else pretrained on someone else's corpus, then continue training its upper layers on your local text. The behaviour shifts. The knowledge underneath does not. You have adjusted the accent of a mind that was formed elsewhere.

Maestro is none of those. The architecture was designed here. The parameters were initialised randomly and trained from that point on our corpus. There is no upstream checkpoint in its history, no foreign model underneath, and nothing to point at and say "well, it started as that one."

The data, and why we can say where all of it came from

The frontier labs have spent three years in court over their training corpora. That is not a problem we wanted to inherit, and it is not a problem a Caribbean institution can afford to inherit on our behalf.

Maestro was trained on publicly available data, with provenance recorded at the document level. No scraped-without-permission corpora. No pirated book collections. No material acquired through a chain that nobody can describe. If somebody asks what Maestro learned from, that question has an answer, and the answer is written down.

This constrained the model. A smaller, cleaner corpus means a smaller model with narrower coverage than one trained on everything anyone could download. We took that trade deliberately. A model that a bank's compliance officer cannot approve is a model that never leaves the lab, whatever it scores on a benchmark.

The corpus is also Caribbean in a way no imported model is. Caribbean English and its creoles, regional institutions, regional law, regional place names, regional history, and the ordinary register people actually write in here. A model that has never seen "the gully", "susu", or "CARICOM Single Market" in context is not going to handle them, and it will handle them wrong with confidence.

World model methods, and what they change

A standard language model is trained on one objective: predict the next token. It gets extraordinarily good at that, and a great deal of apparent reasoning falls out of it for free. It also produces a system whose internal picture of the world is whatever happened to be useful for predicting text.

Maestro's training used world model methods. The training objective extends past next-token prediction toward holding a consistent internal representation of entities, their states, and how those states change when something happens. The model is pushed to keep track of what is true across a passage, not only what word is likely next.

The practical difference shows up in the boring places. A model with a weak world representation will tell you a policy takes effect in January in paragraph two and in March in paragraph five, because both are locally plausible continuations. A model carrying state is more likely to notice it has already committed to January.

We chose this approach for a specific reason. The work this region needs from AI is disproportionately about systems that change over time: a hurricane track, a loan book, a supply chain, a patient. My own research in climate physics builds world models for Caribbean environments and economies so a decision can be simulated before it is taken. Maestro came out of the same line of thinking, applied to language.

Fairness applied during training, not after it

Most deployed fairness work is a filter. The model learns whatever the corpus taught it, including the parts you did not want, and then a second system sits in front of the output and blocks the worst of it. This works about as well as you would expect. The bias is still in the weights. It leaks through paraphrase, through indirect questions, through any prompt the filter's authors did not anticipate.

Maestro applies fairness constraints during training. Representational balance across the groups the model will be asked about was treated as part of the optimisation target, alongside the language modelling loss, with evaluation on Caribbean-specific axes that no imported benchmark covers: nationality within CARICOM, skin tone, creole versus standard English register, rural versus urban, and the informal economy that most Western models read as absence of employment.

This is harder than filtering. It costs compute, it costs corpus engineering, and it constrains the model in ways that show up as lower scores on benchmarks that reward fluent confidence. We did it because the alternative is a model that quietly encodes the assumption that a market vendor in Papine has no income, and no output filter will ever catch that, because the model will never say it out loud. It will just score her lower.

Why it is still in testing

Maestro has not been released. A dedicated red team is actively trying to break it: jailbreaks, prompt injection, extraction of training data, coaxing it into harmful or defamatory output, and the specific failure surfaces a Caribbean deployment would hit, such as impersonation of a public figure or fabricated legal advice.

Holding a finished model back is not caution for its own sake. If the first Caribbean model fails publicly, in front of the governments and institutions it was built for, it will not be read as one team's engineering problem. It will be read as evidence that the region cannot do this, and that verdict will cost the next ten teams their funding. So it stays in testing until we can describe its failure modes, not merely hope we have removed them.

The same reasoning produced TurtleBird, the AI safety toolkit we built and made available free to every government in the Caribbean, for testing whether a model can be pushed into producing harmful output before it goes live rather than after somebody files a report.

What Maestro is not going to do

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

On open-ended reasoning at the hardest end, the large commercial models still lead, and for that work we will keep using them. Maestro's case is different: it covers the region's context properly, its provenance can be audited, its fairness properties were engineered rather than asserted, and nobody outside this region can switch it off.

The weakest part of our own position is scale, and I would rather say that here than have someone discover it in a benchmark table. A model this size will have gaps a bigger one would not. What we have proved is that the floor for building an original model is far lower than this region was told it was, and that the resulting model belongs to the people it serves.

What comes next

• Red-team completion and a public model card, documenting capabilities, known failure modes, evaluation results and the training data policy in full.
• Controlled pilots with institutional partners on narrow, well-specified tasks where the failure cost is contained and measurable.
• Evaluation against Caribbean-specific benchmarks, several of which do not exist yet and which we are building alongside the model.
• Publication of the fairness methodology, so that the next team building a regional model does not have to rediscover it.

Maestro AI Labs was co-founded by Adrian Dunkley and Nicholas Dunkley. The wider portfolio it sits inside, StarApple AI, the Caribbean AI Association, the Caribbean AI Risk Management Council, The Genius Project, and the research programmes behind all of them, is set out at adriandunkley.net/initiatives.html.

Questions people keep asking