New to this? Read it in simple words
- GoTo, an Indonesian tech group, released a new AI model on October 19.
- It understands Indonesian, English, Javanese, Sundanese, Balinese and Batak Toba.
- Anyone can download it for free, under Google’s Gemma licence.
- GoTo says it beats similar models, but it has published no scores.
- Parameters
- The numbers inside an AI model that it learns in training. More usually means a bigger model.
- Weights
- The files that hold a model’s learned numbers. With them, you can run the model yourself.
- Licence
- The rules that say how you may use, change and share something.
GoTo runs Gojek, a ride-hailing and delivery app, and Tokopedia, an online shop. Since 2024 it has released AI models for Indonesian under the name Sahabat-AI. On Monday, October 19, it showed a new one at Bali Tech Week.
What GoTo released
The new model is called GoTo-AI-Gemma-3-27B-IT. It has 27 billion parameters, the numbers a model learns in training, and it can read about 128,000 tokens at once.
It understands Indonesian and English, and four regional languages of Indonesia: Javanese, Sundanese, Balinese and Batak Toba. Its model card names GoTo and AI Singapore, a national research programme, as the developers.
The model is built on Gemma 3, an open model from Google. GoTo merged it with its own models for regional languages, according to the files on Hugging Face.
From the model card and GoTo’s launch statement. The test scores are not published.
Free, with conditions
Startups, businesses, universities, researchers and public bodies can use and adapt it for free, GoTo says. The weights can be downloaded from Hugging Face without a sign-up request.
GoTo calls the model open source. Its licence is Google’s Gemma licence, though, with Google’s own rules for use, not a standard open-source licence. TNGlobal also points this out.
GoTo says the model is small enough to run on servers at universities, startups or public bodies, without a large data centre. At full precision its weights fill about 55 gigabytes, by our count from the model’s files.
What is still GoTo’s claim
In tests against similar open models of its class, GoTo says, the model scored highest on every regional language it tried. The model card lists the tests, IndoMMLU, SEA-HELM, SEA-IFEval and SEA-MTBench, but prints no scores. We found no outside test yet.
GoTo trained it on Indonesian text from media partners and public-sector websites, plus regional-language data, Kontan reports. Universitas Sumatera Utara helped with Batak, and interns from seven universities helped prepare data and test the model.
“Indonesia has made good use of AI built abroad, but we also need to develop our own AI to meet local challenges and needs,” Hans Patuwo, GoTo’s chief executive, said. We translated the quote from Bloomberg Technoz’s report.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
- 3Research · October 19, 2026GoTo Rilis LLM Open-Source, Perkuat Kapabilitas AI Indonesia Bloomberg Technoz
- 4
The model’s details come from its own files. The claim that it beats similar models is GoTo’s alone, from tests whose scores it has not published.
GoTo’s free model for Indonesian and four regional languages opens new building room, though its better-than-rivals claim remains unchecked. How we rate the mood