The Pulse
GoTo Releases 27B Gemma Model for Indonesian Languages
GoTo released a 27-billion-parameter language model adapted for Indonesian, English and four regional languages. The Hugging Face model card lists a 128,000-token context window and requires about 54–60 GB of VRAM for full-precision use.
AI.info Team · Written with AI · fact-checked ·
“Indonesia telah memanfaatkan teknologi AI buatan luar negeri dengan baik. Namun, kita juga perlu mengembangkan AI buatan sendiri untuk menjawab tantangan dan kebutuhan lokal,”
Hans Patuwo, GoTo Group CEO
GoTo released a 27-billion-parameter language model on October 19, making its latest GoTo-AI checkpoint available through Hugging Face. The company announced the release at Bali Tech Week in Sanur, Indonesia. The model, named GoTo-AI-Gemma-3-27B-IT, adapts Google’s Gemma 3 architecture for Indonesian culture and languages, according to its model card.
Six languages, including Batak Toba
The model card lists English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese as supported languages. GoTo says the wider GoTo-AI collection is pretrained and instruction-tuned for Indonesian culture and languages, and credits AI Singapore as a development partner. The release continues GoTo’s open-model work, which began in 2024, according to the announcement.
GoTo says it trained the model on Indonesian-language text from media partners and public-sector websites, along with regional-language data. For Batak, the company worked with the University of North Sumatra to collect material reflecting current usage. The model card also names Indonesian universities, media groups and AI Singapore among collaborators involved in enriching the broader GoTo-AI ecosystem.
GoTo reports strong regional-language results
GoTo says the model earned the highest scores among comparable open models for each regional language tested. The model card describes evaluations using IndoMMLU for local-context questions and SEA-HELM for general language tasks, including translation, question answering and summarization. It also describes instruction-following tests in several languages, with translated and localized evaluation examples.
The card says evaluations used native-language prompts and 100 to 1,000 samples per dataset, depending on the dataset. It does not present the numerical results alongside the model description; instead, it points readers to GoTo’s leaderboard. The reported ranking is therefore the company’s claim, while the card’s methodology gives prospective users details for understanding how its own tests were conducted.
Weights are available under Gemma terms
GoTo describes the release as open source, but the Hugging Face card lists Google’s Gemma license. Google’s terms govern use, modification and redistribution, including restrictions and conditions for sharing copies or derivatives. Developers considering commercial deployment should review those terms rather than treating the downloadable weights as license-free.
Running the 27B model in FP16 or BF16 requires approximately 54–60 GB of video memory, according to the model card. GoTo recommends an 80GB NVIDIA A100, two L40S GPUs or three RTX 4090s; those requirements place full-precision inference beyond many ordinary desktops. The card also warns that the model can hallucinate or give inconsistent reasoning, and advises users to add their own safety tuning and security measures.
GoTo’s release gives Indonesian developers a locally adapted model they can download and run, while its six-language list and evaluation claims set out the company’s intended reach. The practical test for teams will be whether the model performs well on their own regional-language data and workflows—and whether their hardware and use case fit the Gemma license.