Indonesia’s GoTo Group has released a 27-billion-parameter language model trained for Bahasa Indonesia and several regional languages, expanding the company’s effort to build artificial intelligence systems around local linguistic and cultural contexts.
The model, announced on October 19 during Bali Tech Week, is available through GoTo’s official Hugging Face repository. GoTo describes the release as open source, while the public model card lists Google’s Gemma license as the governing license. Organizations considering deployment should review those terms for their intended use.
Six languages and a 128K context window
The instruction-tuned model is based on the Gemma 3 architecture and supports English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese, according to the model card. Its 128,000-token context window is designed to let users process long documents or conversations within a single request.
The repository identifies GoTo as the initiator and GoTo and AI Singapore as the developers. That cross-border development link places the project within a wider Southeast Asian effort to improve language coverage that global models often handle unevenly, although users will still need to test performance on their own terminology, dialects and workflows.
GoTo said the training material included Indonesian-language text from media partners and public-sector websites. For Batak-language data, the company worked with Universitas Sumatera Utara. These descriptions of the training corpus and the model’s local relevance come from GoTo and have not been independently audited.
The company said the model builds on Indonesian open-model work it began in 2024. GoTo and Indosat Ooredoo Hutchison previously introduced the Sahabat-AI initiative, which later included a 70-billion-parameter model and multilingual chat service. The latest release is presented as part of that broader effort, rather than as a direct replacement for every earlier model.
Public weights with substantial hardware needs
GoTo said businesses, researchers and institutions can use, study and adapt the model, including on their own infrastructure. The public model card, however, shows that the full-precision model needs substantial computing resources. It estimates about 54 GB to 60 GB of video memory for FP16 or BF16 inference and suggests hardware such as one Nvidia A100 80GB, two Nvidia L40S cards or three RTX 4090 cards.
That requirement puts the uncompressed model beyond ordinary consumer hardware and means many smaller organizations would need cloud infrastructure, specialized servers or a lower-precision version. The repository provides deployment guidance for inference frameworks including vLLM and Hugging Face Transformers.
The model card lists Indonesian-focused and Southeast Asian evaluation suites, including IndoMMLU and SEA-HELM. GoTo did not provide independently verified performance comparisons in the announcement reviewed by TNGlobal, so the model’s practical accuracy across languages and tasks will need to be assessed by users.
Safety and deployment caveats
The repository cautions that the model can produce inaccurate or misleading information, may struggle with reasoning-heavy tasks and should not be treated as an authoritative source without verification. It also says developers are responsible for safety testing and for applying safeguards suited to their applications.
GoTo said the release is intended to make locally relevant AI more accessible and to support Indonesia’s long-term digital development agenda. The company also framed the model as infrastructure that could help universities, startups and public institutions develop applications without sending all workloads to external platforms.
For enterprises, the practical significance will depend on licensing review, local deployment costs, domain testing and safeguards. The public release gives Indonesian developers another large model to inspect and adapt, but adoption will require more than access to the weights: teams will still need suitable hardware, high-quality application data and systematic evaluation in the languages and use cases they plan to support.
Indosat and GoTo launch new 70 billion parameter Model with multilingual chat service