GoTo releases a 27B open-weights model that speaks Indonesia's regional languages: Javanese, Sundanese, Balinese and Batak Toba

• Who: PT GoTo Gojek Tokopedia Tbk, developed with AI Singapore; announced at Bali Tech Week.
• What: GoTo-AI-Gemma-3-27B-IT — a 27-billion-parameter, instruction-tuned LLM based on the Gemma 3 architecture.
• Languages: 6 — English, Indonesian, Javanese, Sundanese, Batak Toba, Balinese.
• Context window: 128,000 tokens; needs 54–60 GB VRAM (FP16/BF16) — roughly 1× A100 80GB, 2× L40S or 3× RTX 4090.
• When/where: announced October 19, 2026; weights public on Hugging Face under the Gemma license.

Indonesia's GoTo Group has released a 27-billion-parameter language model built for the country's languages — Bahasa Indonesia plus four regional tongues — announced on October 19 during Bali Tech Week. The model, GoTo-AI-Gemma-3-27B-IT, is instruction-tuned on the Gemma 3 architecture and supports six languages: English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese, with a 128,000-token context window.

The weights are public on GoTo's official Hugging Face repository. GoTo describes the release as open source, but the model card lists Google's Gemma license as the governing license — so organizations considering deployment should review those terms for their intended use.

It is not a model you run on a laptop. The model card says loading the full-precision model in FP16 or BF16 needs roughly 54–60 GB of VRAM, and suggests one Nvidia A100 80GB, two L40S cards or three RTX 4090s. The repository ships deployment guidance for vLLM and Hugging Face Transformers, so the intended audience is businesses, researchers and institutions with their own infrastructure.

GoTo says the training material included Indonesian-language text from media partners and public-sector websites, and that it worked with Universitas Sumatera Utara on Batak-language data. The model card credits a longer cast: the University of Indonesia, Gadjah Mada University, Bogor Institute of Agriculture, Bandung Institute of Technology, Udayana University and media groups Kompas Gramedia Group, Republika, Tempo and Hukumonline. For evaluation it uses the Indonesian-focused IndoMMLU and SEA-HELM suites, plus SEA-IFEval and SEA-MTBench for instruction following — but no independently verified performance comparisons were supplied, so real-world accuracy across these languages still has to be assessed by users.

This is not GoTo's first swing at a national model. The release builds on open-model work the company began in 2024 with Indosat Ooredoo Hutchison under the Sahabat-AI initiative, part of Indonesia's digital-sovereignty push — which later included a 70-billion-parameter model and a multilingual chat service. The 27B release is the latest, leaner step in that line: a regional-language model you can inspect, adapt and run yourself, aimed at universities, startups and public institutions that would rather not ship every workload to an external platform.

Sources: TNGlobal, Tech in Asia, GoTo's Hugging Face model card.