GoTo’s downloadable AI model makes room for Indonesia’s languages

An AI assistant becomes more useful when people can ask questions in the language they actually use. GoTo announced a 27-billion-parameter language model at Bali Tech Week on October 19, presenting it as a starting point for locally relevant AI. Its company announcement links directly to the public weights.

The official model card describes a 27-billion-parameter model based on Google’s Gemma 3, developed by GoTo and AI Singapore. It lists six languages: English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese.

For a local business or community organization, the potential value is straightforward: build and test assistance around the words its users choose. Bright sees the release’s significance in that opportunity for local adaptation, rather than in an unverified claim that this model understands every community equally well.

Available weights, with conditions

The repository contains downloadable model weights. These are the learned parameters developers use to run a model, not a ready-made consumer app.

The model card specifies the Gemma license. Gemma’s terms include use restrictions and conditions for redistribution. Downloadable weights should not be presented as an unrestricted open-source license.

There is also a hardware hurdle. GoTo’s card estimates approximately 54–60GB of graphics memory for FP16 or BF16 operation. That is its deployment guidance, not a universal minimum for every possible configuration. A download alone does not make the model practical on an ordinary laptop. Hardware guidance

What still needs testing

The developer describes Indonesian and Southeast Asian benchmark evaluations and warns that the model can invent information. Bright has not independently tested its quality across the six listed languages. Teams should evaluate real questions with fluent speakers before relying on it in a public service.

The company’s October 19 launch announcement and the repository’s history describe different milestones. The repository existed in December 2025, and its visible October 19 changes update documentation. The announcement does not establish that the weights were first uploaded today. Repository history

The useful development is the concrete choice available to builders: a model to inspect and evaluate for local-language services. The human benefit will depend on what those services can reliably do for the people using them.