Skip to content
AI

GoTo releases 27-billion-parameter model co-developed with AI Singapore

The Indonesian tech group partnered with AI Singapore to release a regional language model built on Google's Gemma 3 architecture.

OSH Desk1 min read

GoTo Group announced a 27-billion-parameter language model co-developed alongside AI Singapore during Bali Tech Week, aiming to enhance generative AI capabilities for Southeast Asian languages [1]. The architecture is built on Google's Gemma 3 base and handles a context window of 128,000 tokens across six supported languages [1].

Publicly released through Hugging Face, the model makes its weights available to developers under Google's Gemma license rather than an open-source standard [1]. Benchmarking for the release relied on the IndoMMLU and SEA-HELM evaluation frameworks [1]. Its model card notes that the system may generate inaccurate outputs under certain conditions [1].

Running the model at full precision requires 54GB to 60GB of video memory, making an Nvidia A100 80GB or comparable accelerator the recommended setup for hosting the system [1].

Questions and answers

What architecture is the GoTo model based on?

The 27-billion-parameter system is built using Google's Gemma 3 architecture and operates with a 128,000-token context window.

Who helped build the model?

GoTo Group co-developed the model in partnership with AI Singapore.

What hardware is needed to run the full-precision weights?

The model requires 54GB to 60GB of VRAM, with recommended setups including Nvidia A100 80GB accelerators or equivalent GPUs.

TopicsOpen-source AIFoundation models

Companies & investors in this story

Sources

  1. 1.GoTo’s 27B model speaks Batak, but the real test is who can afford to run it — e27, Oct 20, 2026

Researched and drafted with AI tools under OSH editorial rules. Every fact links to its source. Editorial policy · Suggest a correction