20 OCT 2026 — GoTo, the Indonesian group behind Gojek and Tokopedia, has put a 27-billion-parameter language model trained for Bahasa Indonesia and four regional languages on Hugging Face for anyone to download. The company announced the release at Bali Tech Week on 19 October 2026. It is pitching the model at universities, startups, businesses and public institutions that want to build their own AI tools without starting from nothing.
GoTo's earlier open models came out under the Sahabat-AI brand, which it launched with the telecoms group Indosat Ooredoo Hutchison. The new model carries GoTo's own label, and its model card does not mention Indosat. It does credit Singapore's national AI programme as a co-developer.
What GoTo released on Monday
The press release lists the languages as Bahasa Indonesia, Javanese, Sundanese, Balinese and Bataknese. The model card names the model GoTo-AI-Gemma-3-27B-IT, adds English, and gives the Batak variety as Batak Toba. It says the model uses Google's Gemma 3 architecture and tokenizer and handles a context of 128,000 tokens.
According to GoTo, the training data came from Indonesian-language text supplied by media partners, from public-sector websites and from regional-language sources. For Batak, the company says it worked with Universitas Sumatera Utara to gather data on how the language is spoken now. Interns from seven universities, Udayana University in Bali among them, worked on data preparation, training and testing.
The case GoTo is making
Group chief executive Hans Patuwo framed the release as a national contribution. "Indonesia makes great use of global AI, but we must build our own too, to solve the problems that are unique to us," he said in the company's statement.
He argued that global models may handle local data and home languages poorly. GoTo says it already builds this kind of AI for its own customers. Opening the model, he said, gives other Indonesian builders "a starting point rooted in the local context". The company also linked the release to the Indonesia Digital Vision 2045 set out by the Ministry of Communication and Digital Affairs (Komdigi).
A Singapore team behind much of it
The press release does not mention Singapore. The model card does. It lists PT GoTo Gojek Tokopedia Tbk and AI Singapore as developers, and its team section names six GoTo staff and 35 people from AI Singapore. AI Singapore is a government-funded programme based at the National University of Singapore, with backing from the National Research Foundation, according to the card's acknowledgements.
AI Singapore has worked on GoTo's models before. The company's November 2024 release said Sahabat-AI was supported by AI Singapore and Tech Mahindra, and launched with 8- and 9-billion-parameter models. A June 2025 follow-up added a 70-billion-parameter version and a chat service, and said the first models had been downloaded more than 35,000 times. RECATOOLS lists both Sahabat-AI and AI Singapore's own SEA-LION models in its AI directory.
What "open-source" means here
GoTo's headline calls the model open-source. The card puts it under the Gemma licence, which is Google's own terms of use and not a standard open-source licence. Those terms let users reproduce, modify and distribute the model, but anyone passing it on must carry Google's use restrictions with it. Uses barred by Google's prohibited-use policy are forbidden. The terms also say Google "reserves the right to restrict (remotely or otherwise)" use it believes breaches the agreement.
None of that stops a university or a startup from using the model. A company planning a commercial product on it should read Google's terms rather than rely on the word in the headline.
Running it is not trivial
GoTo describes 27 billion parameters as mid-sized, small enough to run on an institution's own servers. The card puts the requirement at roughly 54 to 60 GB of GPU memory at 16-bit precision. It suggests one 80 GB Nvidia A100, two L40S cards or three RTX 4090s. That is within reach for a university lab. It is not a laptop.
The sample code on the card points to a repository path, GoToCompany/GoTo-AI-Gemma-3-27B-IT, that returned an authorisation error when RECATOOLS checked it on 20 October. A comment in the code tells users to substitute the real path. The working address is the one GoTo's press release gives.
What is still missing
The card describes four benchmarks, IndoMMLU, SEA-HELM, SEA-IFEval and SEA-MTBench, but prints no scores. It points readers to a separate GoTo leaderboard instead. The Indonesian outlet youngster.id reported that the model posted the top score for regional-language processing on comparable benchmarks. The claim does not appear in GoTo's English release, and RECATOOLS has not seen published results that confirm it.
The card also states the usual caveats: the model can hallucinate, and developers are told to do their own safety tuning. The real test is whether Indonesian universities and public bodies adopt it, and whether GoTo publishes numbers that let outsiders check the claims.
View author profile → · Editorial policy
