GoTo Release Model 27 Billion Parameters, Language

GoTo released a new model of artificial intelligence with 27 billion parameters capable of understanding Indonesian language and a number of local languages, including Java, Sunda, Bali and Batak.

The AI model is available for free for startup, companies, universities, researchers, to public institutions. They can use and develop these technologies to create AI services that are more suited to the need of Indonesian society.

This major language model or large language model was built using Gemma 3 architecture. Based on documentation on Hugging Face, it supports six languages, which are English, Indonesia, Java, Sunda, Bali and Babak Toba.

The CEO of the GoTo Hans Patuwo Group says Indonesia needs to develop its own AI technology to not only rely on foreign products.

"Indonesia has utilized the technology of the global AI well. However, we also have to build our own AI to solve typical problems for us," Hans said in GoTo's official statement.

According to Hans, the AI developed for global needs is not necessarily able to understand Indonesian language and habits optimal, especially the use of regional language in everyday life.

Therefore, GoTo wants to provide the foundation of AI that understands the local context and opens the opportunity for Indonesian developers to create different artificial intelligence-based services.

In its development, the AI model was trained to use Indonesian language text from several media partners and public sector websites. GoTo also took the University of Northern Sumatra (USU) to collect data of Batak that reflects the use of that language today.

Technically, this model has a context capacity of up to 128,000 tokens, allowing AI to process long documents or conversations with a lot of information in one request.

With 27 billion parameters, that model goes in the middle category. Parameters are a value that AI learned during training to recognize patterns and generate responses, even though the number of parameters does not automatically determine the level of accuracy.

GoTo called this model to be able to run through university's infrastructure servers, startup, and public institutions without having to build large-scale data centers.

Even so, the computing needs are pretty high. Model documentation includes GPU memory needs about 54-60 GB to run it on the precision of FP16 or BF16, with Nvidia A100 80 GB as one recommended configuration.

GoTo's latest AI model is accessible via the Hugging Face platform here.

GoTo called this technology an open source model. However, the documentation includes Gemma's license, so the developers still need to pay attention to the terms of their use and distribution.

Like any other AI generative model, this technology also has potential wrong answers or hallucinations. Therefore, developers need to do a test before integrate it into public service.

The launch of this 27 billion parameter continued the development of an open AI which has started GoTo since 2024. The company also involves interns from seven universities in Indonesia in data gathering, training, to Al's testing models.