Eigen RadarAI
Archive
Analysis

GoTo opens a 27 billion parameter model for Indonesian and regional languages

GoTo has released downloadable weights for a model focused on Indonesia’s languages and cultural contexts. Developed with AI Singapore on Gemma 3 architecture, it supports Indonesian and several regional languages alongside English. The release extends work begun in 2024, with local media and public-sector text among the training material. Adaptation remains subject to the Gemma licence and deployment-specific safety work.

Artificial Intelligence··Evening
In an imagined machine world, a mesh matrix extends on rails from a larger shared structure among tropical stone terraces.

GoTo releases weights for local-language use

GoTo, an Indonesian technology company, released a 27 billion parameter language model at Bali Tech Week, a technology event in Denpasar, Indonesia, on 19 October. The release focuses on Indonesian and regional languages. The weights can be downloaded for adaptation rather than accessed only through a hosted service.[1], [2]

The model was developed by GoTo and AI Singapore, its Singapore-based AI development partner, using Gemma 3 architecture. The instruction-tuned version is distributed through Hugging Face, a platform for sharing machine-learning models. Its model card lists 6 languages: English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese.[1]

Local text supports the language focus

GoTo says its training material includes text from local media partners and public-sector websites. It worked with Universitas Sumatera Utara, a university in Indonesia, on Batak-language data. The project continues local model work begun in 2024 and emphasizes Indonesian linguistic and cultural contexts. The card specifies a 128000-token context window, the space available for processing input and surrounding text.[1]

Adaptation remains subject to licence and safety requirements

Universities, researchers, startups and public institutions can download and adapt the model under the Gemma licence. Developers recommend further safety fine-tuning and security measures for specific deployments. They also warn of hallucinations and inconsistent reasoning. The release includes a Transformers usage example, giving developers a software route to load the instruction-tuned weights.[1]

References

  1. News sourceBali TribuneGoTo releases weights for a 27-billion-parameter model for Indonesian languages↩1↩2↩3↩4
  2. News sourceSuara.comGoTo opens its 27-billion-parameter model for local languages↩
AI safety and regulationThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →