Products / The Papiamentu modelsCuraçao

A shop owner in Punda reading a WhatsApp message on her phone behind the counter, the colourful street outside.

The Papiamentu models

Models that speak Papiamentu.

Chuchubi writes it, Orea hears it and Kanto is learning to say it. Our own models, trained for the Curaçao variety.

Chuchubi · ShippedOrea · ShippedKanto · In trainingThe app · In development

Try Chuchubi Mini free Book your assessment

Mini is free once you sign in with an email link. No password and no card. Read the evals

01Papiamentu, Dutch, English and Spanish on one model
02Spelling checked against the official Curaçao spelling (the FPI standard)
03Speech to text that follows the switching
04Measurements published at Galvan Intelligence

Four products, one language first.

We train our own models for Papiamentu, Curaçao first.

A business on Curaçao quotes in Dutch, sells in Papiamentu, reports in English and answers suppliers in Spanish, often in the same afternoon.

General AI models handle Papiamentu poorly. On test sentences the models never saw in training, a leading general model had spelling errors flagged in most of its replies. Our models are trained for the Curaçao variety and checked against its official spelling.

Because the models are ours, we can run them for you, on a setup reserved for you, or on your own machines for larger projects.

Chuchubi V1.1.1 Mini● checker on

Mi ta yama pa puntra si mi pedido ta yega awe.clean

Te ki ora boso ta habrí awe tardi?clean

Mi ke traha un sita pa otro siman.clean

Written by Chuchubi V1.1.1 Mini in Chat Studio on 2026-10-22, used as written.

ChuchubiReads and writes Papiamentu, and checks every reply against the official spelling.Shipped
OreaSpeech to text, including the switching between languages that is how people talk. More about OreaShipped
KantoText to speech in a Curaçao voice. Nothing to try yet.In training
The appOne app that listens, answers and writes in Papiamentu, built on all three.In development

Chuchubi at work.

Real screens from Chat Studio, captured on 23 October 2026. The Papiamentu is exactly what Chuchubi wrote.

How Chuchubi learns, and checks itself.

We start from an open model and teach it Papiamentu. That teaching is what we sell. This is the path from published text to a reply you can trust on spelling.

Step 01 of 05

Real text first

Published text supplies the vocabulary.

So The words Chuchubi uses are words people on Curaçao write.

Step 02 of 05

Teach the base the language

Extra reading in Papiamentu teaches the base model the language first.

So Task training shapes how it answers. This stage is where it learns the language.

Step 03 of 05

Then the tasks

Worked examples, generated and then checked.

So Scores are measured on prompts the model was not trained on.

Step 04 of 05

Every reply is checked

A dictionary and the FPI standard.

So On sentences it never saw in training, most replies came back with nothing flagged.

Step 05 of 05

You use it

Mini free with sign-in. Standard and Max per account.

So You can test it on your own sentences today, without a card.

What the checker is for

It reads the Papiamentu, never the facts.

The checker tells you whether a reply is spelled the way the standard spells it. Whether the content is true is a separate question, and the checker does not answer it. Full measurements are at Galvan Intelligence.

Each model, with its real status.

Two you can use today, one in training and one in development.

Chat Studio: a Papiamentu reply from Chuchubi to a translation request, marked clean by the spelling checker, above a second reply with one word flagged as not in the dictionary.

TierMiniFree with sign-in

TierStandardPer account

TierMaxPer account, on request

Three tiers of Chuchubi V1.1.1. The test set cannot yet tell them apart on errors, so they are not ranked against each other.

TextShipped

Chuchubi

Reads and writes Papiamentu in the Curaçao variety, and handles Dutch, English and Spanish on the same core. Every reply is checked against the official spelling. Mini is free once you sign in; Standard and Max are opened per account.

Curaçao variety
Spelling checked
Four languages
Evals published

Try Mini free Read the evals

A woman sitting on a wall in Otrobanda at dusk, speaking Papiamentu into her phone.

Speech to textShipped · V1.12

Orea

Turns spoken Papiamentu into text, including the switch to Dutch or English in the middle of a sentence. Orea Studio takes long recordings with several speakers, times every line and marks the words it is unsure of.

Code-switching
Speakers separated
SRT, VTT, text, JSON or Word
Apache 2.0

More about Orea Open Orea Studio

Text to speechIn training

Kanto

Will speak written Papiamentu in a Curaçao voice. The hard part is the material: there is almost no recorded Papiamentu built for this. Nothing to try yet, and no numbers published.

Listens, answers, writesIn development

The app

One app that listens, answers and writes in Papiamentu, built on Chuchubi, Orea and Kanto. It is in development, and no release date has been announced.

For work that happens in Papiamentu.

Concrete uses for the two shipped models, today.

Chuchubi

Customer messages

Draft replies to customers in the language they wrote in, spelled the way the official standard spells it.

Orea

Meetings and interviews

Turn a recording into a timed transcript with the speakers named, then export it.

Private setups

Sensitive data

Run the models on a setup reserved for you, or on your own machines for larger projects, so the text stays close.

Builders

Your own product

Put Papiamentu into a system of your own: Standard and Max are opened per account after we read what you are building.

What the models do not do.

Published with the measurements, because a measurement is only useful with its limits.

The checker does not check facts.
It tells you whether a reply is spelled correctly. It says nothing about whether the content is true.
It is Curaçao Papiamentu, not Aruban Papiamento.
Aruban Papiamento is filtered out and not supported.
Small differences cannot be measured yet.
Our test set is still too small to show which tier makes fewer mistakes, so we do not rank them.
The test leans formal.
The evaluation data leaned formal, so the scores may not reflect everyday messaging.
Conversations have a length limit.
Context is finite, and older turns are dropped automatically.
Kanto and the app are in development.
There is nothing to try for either yet, and no numbers.