GoTo’s Regional-Language AI: What Is Actually New?

In a village in Bali, a child listens to their grandmother tell stories about their ancestors in a language that never appears in school textbooks. The child understands every expression, including subtle irony that cannot be translated word for word.

When the child starts school, they learn to use artificial intelligence to do assignments. The AI responds in polished Indonesian, but suddenly loses its intelligence when the grandmother slips in a Balinese expression.

The machine has mastered grammar, but may not understand how people live. Imagine if one day it could follow their conversation without first asking everyone to learn to speak like people in Jakarta.

That aspiration is what made GoTo’s announcement on 19 October 2026 noteworthy. The technology company behind Gojek and Tokopedia launched a large language model (LLM) with 27 billion parameters.

The model is designed to understand Indonesian as well as Javanese, Sundanese, Balinese and Batak. GoTo has made it available so that start-ups, universities, researchers and public institutions can develop AI applications suited to Indonesia’s needs.

But before hailing it as the birth of Indonesian artificial intelligence, we should examine a fundamental question: what is genuinely new about GoTo’s work? Has the company invented new AI technology with regional-language capabilities?

The question matters if we are to avoid getting caught up in launch-day fanfare. The technology world is fond of presenting new products on grand stages, while the real innovation often takes place in quiet workspaces.

There are lengthy processes that the public does not fully see: collecting data, cleaning text, training the model, testing its responses, and repeating the same work until the machine understands people better.

An important clue is contained in the name of the AI model GoTo uploaded to Hugging Face: “gemma3-27b-cpt-gotoai-v3-instruct”. This technical name reveals its lineage. GoTo used Google’s Gemma3-27B and then adapted it to the Indonesian context.

CPT refers to continued pre-training: further training a model that already has language capabilities. The model can then undergo instruction tuning.

This final process helps the AI follow instructions and answer questions more effectively. GoTo, therefore, has not announced the invention of a new LLM architecture. It has developed the capabilities of an existing model through training and adaptation based on local data.

That strategy makes sense. Building a large language model from scratch requires enormous quantities of data, costly computing power, specialist expertise and considerable time.

Developing an existing foundation model allows a team to devote more resources to improving the capabilities it considers important.

In GoTo’s case, those priorities include Indonesian, regional languages and local cultural context. So, what is new?

First, GoTo has taken its model development to a third version, as indicated by the V3 in its name. This points to ongoing development, not a project that began in October 2026.

Second, the company has released a 27-billion-parameter model specifically geared towards Indonesian and four regional languages. Third, GoTo has opened access to the model so that others can study, use and adapt it.

The third point may be the most important. Indonesia already has many people capable of building applications, but not all of them have the resources to train a large language model.

By providing downloadable model weights, GoTo gives developers a starting point for creating education services, translation tools, cultural documentation and public-service applications.

Before calling the release an unprecedented breakthrough, we should look at similar work already under way.

AI Singapore has developed SEA-LION (Southeast Asian Languages in One Network), a family of language models focused on the linguistic and cultural needs of South-east Asia. The programme has developed models since its early generations, including models based on Gemma and Llama.

The idea of using global foundation models to strengthen regional languages, then, was already gaining ground before GoTo’s latest announcement. SEA-LION also shows that this work can encompass languages across South-east Asia, not just Indonesian and four regional languages.

GoTo itself has a history of developing local AI through its collaboration with Indosat in the Sahabat-AI ecosystem. That initiative has already offered models aimed at Indonesian and several regional languages, including Javanese, Sundanese, Balinese and Batak.

For that reason, the October 2026 release is better understood as continuing and strengthening GoTo’s existing capabilities, rather than marking the birth of Indonesia’s first regional-language AI.

Changes in model size, version, data coverage and public access can still represent progress. But that progress should be judged by results that can be scrutinised, not simply by the launch date.