For much of the generative AI boom, a user in North Sumatra typing a query in Batak Toba has been met with a machine that, at best, partially understands them. GoTo Group aims to change that dynamic, though the most difficult challenges are just emerging.
Indonesia’s largest digital ecosystem unveiled a 27-billion-parameter language model trained for Bahasa Indonesia and several regional languages on Monday at Bali Tech Week, publishing the model weights to its official Hugging Face repository.
Built on Google’s Gemma 3 architecture, the instruction-tuned model supports six languages: English, Indonesian, Javanese, Sundanese, Batak Toba, and Balinese. It also features a 128,000-token context window, allowing it to process lengthy documents or extended customer-service conversations in a single pass rather than in fragmented segments.
The repository lists GoTo as the initiator, with GoTo and AI Singapore credited as co-developers. This cross-border collaboration situates the release within a broader regional effort—anchored repeatedly by AI Singapore—to address a gap global models have long overlooked: Southeast Asia’s languages are spoken by hundreds of millions, yet remain marginalized in most training datasets.
“Open Source” with Qualifications
GoTo describes the model as open source. However, the model card reveals a nuance: the governing license is Google’s Gemma license, not a standard open-source license such as Apache 2.0 or MIT. While Gemma’s terms allow broad use and modification, they include Google’s own usage restrictions.
For a hobbyist, the distinction is largely academic. For a bank, telco, or government agency planning to embed the model into a product, it is the first item a legal team will scrutinize.
This is not unique to GoTo. Much of what the industry markets as “open” AI is more accurately “open weights”: the parameters are downloadable, but the rules are set by someone else.
Nevertheless, for a release explicitly framed around reducing dependence on foreign platforms, building on a license controlled by a U.S. tech giant presents an irony worth noting. It sharpens a question regional observers have raised for months: whether Southeast Asia can achieve sovereign AI on borrowed foundations.
From Sahabat-AI to a Leaner Approach
This is not GoTo’s first venture into homegrown language models. In 2024, the company and Indosat Ooredoo Hutchison launched the Sahabat-AI initiative, which produced a 70-billion-parameter model and a multilingual chat service. GoTo positions the new model as a continuation of that effort rather than a replacement.
The shift to a smaller 27-billion-parameter model appears pragmatic. Larger is not necessarily better when the target users are Indonesian universities and startups rather than hyperscalers. The timing is also significant. GoTo recently posted its first quarterly net profit after years of restructuring and subsidy reductions, and leadership under CEO Hans Patuwo has strong incentive to demonstrate to investors that AI is a strategic asset, not merely a cost center.
Regarding training data, GoTo states it utilized Indonesian-language text from media partners and public-sector websites, and collaborated with Universitas Sumatera Utara on Batak-language material. None of this has been independently audited. That gap extends beyond accuracy. As newsrooms globally debate the terms on which their journalism should feed AI systems, the specifics of GoTo’s arrangements with its media partners warrant greater transparency than a single line in an announcement.
The 60GB Barrier
Here the “AI for everyone” narrative meets the reality of the server room. According to the model card, running the full-precision model requires approximately 54GB to 60GB of video memory. The suggested hardware includes one Nvidia A100 80GB card, two Nvidia L40S cards, or three RTX 4090s.
In practical terms, this is not a model a two-person startup in Yogyakarta can run on a gaming laptop. Most smaller organizations will require cloud infrastructure, specialized servers, or a compressed, lower-precision version that sacrifices some quality for affordability. GoTo’s repository provides deployment guidance for frameworks such as vLLM and Hugging Face Transformers, which assists implementation but does not reduce the underlying cost.
This makes Indonesia’s compute strategy part of the story. Jakarta has been courting Nvidia and AWS to expand domestic chip and cloud capacity. A local-language model is only as sovereign as the infrastructure it runs on; if most deployments ultimately reside on foreign-owned clouds, the independence argument weakens considerably.
Benchmarks Without Independent Verification
The model card references Indonesian and Southeast Asian evaluation suites, including IndoMMLU and SEA-HELM. However, GoTo has not published independently verified comparisons against rival models, so claims of superior local-language performance remain unsubstantiated. The repository itself is candid: it warns that the model can generate inaccurate or misleading information, may struggle with reasoning-heavy tasks, and should not be treated as authoritative without verification. Safety testing and safeguards, it adds, are the developer’s responsibility.
A Crowded Field of Regional Contenders
GoTo enters a competitive arena. AI Singapore’s SEA-LION family has long targeted regional languages, Thailand’s SCB 10X has advanced its Typhoon models for Thai, and Malaysia has made local-context AI a sovereign priority.
Above them all sit global giants—Google, Meta, Alibaba—whose general-purpose models continue to improve in Southeast Asian languages, driven partly by rising regional usage in Vietnamese, Thai, and Indonesian. GoTo’s potential edge lies in lower-resource languages such as Batak Toba and Balinese, where global players have little commercial incentive to invest.
The Bottom Line
GoTo frames the release as infrastructure for Indonesia’s long-term digital agenda—a foundation universities, startups, and public institutions can build on without exporting every workload abroad. The ambition is commendable, particularly for languages the AI industry has historically ignored.
But access to weights is not adoption. Teams will still need a licensing review, suitable hardware, clean application data, and rigorous testing in the dialects and domains they actually serve. GoTo has opened the door. Whether Indonesia’s developers can afford to walk through it remains the real test.
