Two Sovereign Language Models Shipped 17 Days Apart. The Caribbean's AI Roadmap Still Names None.

I spent part of August lining up the Caribbean's new AI roadmap against what other regions with underserved languages have actually shipped this year. Three of its four pillars already have something built against them: a governance report exists, upskilling pilots are running at UWI, and resilience work is tied to the region's hurricane-prediction efforts. The fourth pillar, Culture and Creativity, does not. It names the right problem and stops there.

A Roadmap With a Pillar and No Model

On 7 July 2026, at its 126th Special Meeting, CARICOM's Council for Trade and Economic Development configuration on Information and Communications Technology formally endorsed the UNESCO Caribbean Artificial Intelligence Policy Roadmap. Per UNESCO's own account and the UN's Caribbean office, it was the first time a CARICOM principal organ had collectively endorsed an AI policy document, built from consultations with more than 1,000 institutions and individuals across 20 English- and Dutch-speaking territories.

The roadmap rests on four pillars: Culture and Creativity, Governance and Transformation, Education and Upskilling, and Resiliency and Sustainability. The Culture and Creativity pillar's stated aim is to protect the region's linguistic and cultural identity as AI tools spread, and its text calls specifically for integrating Caribbean dialects, Patois and Creole into AI-driven chatbots and educational outreach, alongside more funding for regional-language processing in music, heritage tourism and the wider creative economy.

What the pillar does not contain is anything a technical team could execute against. It names no model architecture, no target language list, no committed corpus, no budget line and no delivery date. The three other pillars each already point at a concrete instrument: a finished governance report from the Caribbean AI Task Force, upskilling work running through UWI's new I-INSIGHT institute, and resilience modelling tied to the region's climate-prediction efforts. Culture and Creativity is the one pillar still waiting for its instrument.

Research compiled by StarApple Analytics puts Caribbean adult use of generative AI at roughly 13%, against about 55% worldwide, and every tool behind that 55% figure was trained overwhelmingly on standard English. A Haitian Creole speaker typing into a chatbot trained on Parisian French and American English is using a different tool from the English speaker's, one that only shares its interface.

India Trained Two Models From Scratch for 23 Languages

Sarvam AI gives a working answer to what an "instrument" for a pillar like this would actually look like. On 6 March 2026, the Indian company open-sourced two large language models: Sarvam 30B, a 30-billion-parameter dense model, and Sarvam 105B, a mixture-of-experts model with 105 billion total parameters and 10.3 billion active per request. Both were trained from scratch, not fine-tuned on top of an existing foundation model, using datasets built around the 22 languages recognised under the Eighth Schedule of the Indian Constitution, plus English, for 23 languages in total.

Sarvam released the weights under the Apache 2.0 licence, on Hugging Face and on AIKosh, the Indian government's own AI data and model repository. Hosting on a government-run repository alongside the standard global one was a deliberate choice: it positions the release as public infrastructure the state can point to, not only a startup's open-source moment.

Sarvam's models sit on top of Bhashini, India's National Language Translation Mission, a multi-year government programme that spent years building parallel text and speech corpora across Indian languages before any lab attempted to train a model of this size on top of them. The model was the last step in that sequence.

A Model Trained on 64 GPUs, Not a Data Centre

Sarvam's release makes the point that a serious multilingual model no longer has to come out of Silicon Valley. Cohere Labs, the research arm of the Canadian AI company Cohere, shows it does not need a data centre budget either. On 17 February 2026, on the sidelines of the India AI Summit, Cohere Labs released Tiny Aya, a family of five models built around a 3.35-billion-parameter pretrained base, tuned to cover more than 70 languages, many of them low-resource.

Three of the five Tiny Aya models are regional variants: Tiny Aya Fire for South Asian languages, Tiny Aya Water for Asia-Pacific, West Asia and Europe, and Tiny Aya Earth, tuned specifically for African languages including Swahili, with attention to culturally specific names and terms. On mathematical reasoning benchmarks run in African languages, Tiny Aya Earth scored 39.2%, against 17.6% for Google's Gemma 3 4B, a general-purpose model of roughly the same size that was not built with those languages as a priority.

The training budget is the number that matters most in this release: 64 Nvidia H100 GPUs, run for long enough to produce a model that works offline on a consumer laptop. That sits far below a hyperscaler's compute bill and closer to what a well-funded university lab, a single company, or a modest regional government contract could book without a headline capital raise attached to it.

The Caribbean's Bottleneck Is the Corpus, Not the Compute

Compute no longer explains the gap. Tiny Aya's 64-GPU training run and Sarvam's Apache-licensed weights both show that a serious language model for an underserved population is now a project a regional government, a university consortium or a single well-capitalised company could fund without waiting on a hyperscaler's goodwill. What neither Sarvam nor Cohere had to build from zero was the underlying text. India had Bhashini's years of translation and speech corpora already assembled. Pan-African NLP research collectives such as Masakhane have spent a decade building parallel corpora across dozens of African languages, the exact groundwork a project like Tiny Aya Earth depends on existing somewhere before any lab can train against it.

Haitian Creole has roughly 12 million native speakers, more than several of the individual languages Sarvam trained for. Jamaican Patois is the daily language of an estimated 3 million people. Trinidadian Creole, Bajan and a dozen smaller Caribbean Creoles add several million more. None of that population is small by the standard Sarvam and Cohere worked against. What is missing is the standardised, rights-cleared, digitised corpus of Creole text and speech a training run would draw on, the layer Bhashini and Masakhane spent years building before either region tried to put a model on top of it.

The Caribbean AI Task Force's own final report, launched at the first Caribbean AI Forum in Trinidad in July 2026, argues the region's more realistic task is not to build frontier models but to set the terms under which external models, platforms and standards enter it. That is sound advice for something the size of Sarvam 105B. It is the wrong advice for something the size of Tiny Aya, which was never a frontier bet in the first place. A three-billion-parameter Creole model trained on a purpose-built corpus is a different kind of project to a 105-billion-parameter reasoning model, and treating the two as the same ambition is a large part of why Culture and Creativity has nothing built against it yet.

Where the corpus lives is a sovereignty question in its own right, not only a technical one. Jamaica's Data Protection Act carries fines of up to 4% of global turnover for mishandled personal data, and a Creole speech corpus built from real conversations, voice notes and interviews is personal data by any reasonable reading of that law. A regional Creole effort would need the same data-localisation discipline the region's banks and insurers are only now building for financial records: informed consent, a named data controller, and infrastructure that keeps raw recordings inside a jurisdiction that can be held to account for them. None of the three pillars with an existing instrument, governance, upskilling or resilience, had to solve that problem first. Culture and Creativity does.

Domain-specific Caribbean products already show the pattern a Creole model would need to follow. SportsBrain AI, the region's sports intelligence platform, works because it was built on match, player and league data that most global sports-analytics tools never collect for Caribbean competitions specifically. A Creole language model needs the same asymmetry: a corpus that exists because someone in the region deliberately built it, not because a global lab's web crawl happened to catch a few thousand Patois posts on its way past.

Maestro AI Labs' own approach, retraining existing open-weight foundation models on approved Caribbean data rather than training from zero, is one live attempt at the corpus side of this problem, and it runs into the same constraint every other regional effort does: the training data has to be collected, cleared and localised before a single GPU hour gets spent on the model itself. That work has not produced a public Creole benchmark yet. Neither has anyone else in the region. Whether a retrained open-weight model or a from-scratch small model like Tiny Aya performs better on a language with almost no digitised text to begin with is still an open question nobody has tested for a Caribbean Creole, because nobody has assembled the corpus a test would need.

What a Caribbean Attempt Would Actually Require

The roadmap does not put a date on the Culture and Creativity pillar. The other three pillars in the same document each point at something with a deadline attached: a report that was due and delivered, a hub that is rolling out campus by campus, a resilience system already in testing. A pillar with no date behind it is a pillar nobody has yet been assigned to deliver.

A narrower ambition would change that: pick one Creole, most likely Jamaican Patois or Haitian Creole given the size of population each commands, and fund the groundwork first, a standardised orthography, a rights-cleared corpus of text and transcribed speech, and a benchmark test set that a Sarvam-style or Tiny-Aya-style training run could then be measured against. That is a multi-year, largely invisible project with no press release attached to its early stages. It is also the only path any of the three regions covered here actually took before a model shipped.

The next COTED-ICT meeting will show whether the Caribbean takes that path: either Culture and Creativity appears on the agenda with a budget line attached, or it is still described as an aspiration a year after every government signed it. The other three pillars moved from document to working instrument inside twelve months, so the open decision is whose corpus gets built first.