GoTo makes its Indonesian-language AI available to the public

[SINGAPORE] Gojek operator GoTo has released a 27-billion-parameter artificial intelligence model tailored for Indonesia, which it says outperforms an earlier model more than twice its size while requiring less computing power.

Launched at Bali Tech Week on Monday (Oct 19), the model supports Bahasa Indonesia and four regional languages: Javanese, Sundanese, Balinese and Batak.

GoTo is making the model available for free on AI platform Hugging Face, allowing universities, startups, businesses and public institutions to use and customise it for their own needs, including commercial applications.

GoTo group chief data officer Ofir Shalev told The Business Times that the model was tested against similarly sized open models and the company’s earlier 70-billion-parameter model.

It achieved the highest overall score across the five Indonesian languages, leading in all four regional languages. It also outperformed GoTo’s earlier model despite having less than half as many parameters, he said.

The gains come from training the smaller model more extensively on Indonesian language data, rather than simply increasing its size.

GoTo trained the model on about 65 billion tokens of text, including around 40 billion in Bahasa Indonesia and 16 billion in English. A further 10 billion came from the four regional languages.

To build the dataset, GoTo worked with universities, government agencies and media organisations to gather Indonesian and regional language text. It also partnered Universitas Sumatra Utara to collect contemporary Batak language data, with efforts to expand the dataset ongoing.

Rather than develop the model from scratch, GoTo built on Google’s open Gemma 3 model. It spent five days carrying out further training using 64 Nvidia H200 graphics processing units, followed by three days refining the model with instruction data developed with local universities, Shalev said.

SEE ALSO

‘Not trying to solve world hunger’: GoTo takes pragmatic approach to AI

The process was designed to improve its understanding of local languages and cultural context, while making its responses sound more natural to Indonesian users.

Lower computing requirements

GoTo is positioning the smaller model as an option for organisations seeking to develop AI applications without investing in extensive computing infrastructure.

At full precision, it requires about 55 to 65 gigabytes of GPU memory, allowing it to run on a single Nvidia A100 80 GB data centre GPU, or two to three high-end workstation graphics cards, Shalev said.

A compressed version can run on a single high-end consumer graphics card, albeit with some loss of accuracy. Running costs would depend on the scale and nature of deployment.

GoTo has also deployed the technology within its own services.

For instance, the model has been used to help merchants joining GoFood by reading photographs of printed menus and automatically generating menu listings. In customer support, it has been used to assess complaints supported by photographs, including identifying missing items in food orders and verifying proof of delivery.

Shalev said newer, more advanced models now power some of GoTo’s key services, including its Dira Help and Dira Voice assistants, as well as food search and recommendations.

The company may eventually release some of these models publicly, giving developers access to more advanced tools.

GoTo group CEO Hans Patuwo said developing AI trained on Indonesian languages and data was important because global models might not fully capture the country’s local context.

“Indonesia makes great use of global AI, but we must build our own too, to solve the problems that are unique to us,” he said.

“Opening this technology to everyone is our way of giving back to the country, providing anyone building AI in Indonesia with a starting point rooted in the local context.”