GoTo releases a 27B Gemma model for Indonesian languages

GoTo and AI Singapore released downloadable 27B weights tuned for six languages. The model card lists the Gemma license, a 128K context window, demanding VRAM needs and publisher-run evaluations.

2 min read

Busy Jakarta street with pedestrians and traffic; illustrative image for a locally adapted Indonesian-language AI model, not a model screenshot

GoTo released GoTo-AI-Gemma-3-27B-IT on October 19, a downloadable 27-billion-parameter model developed with AI Singapore and tuned for Indonesian languages and context. The release adds a locally focused option to Google’s Gemma 3 family rather than introducing a new base architecture.

GoTo says the model was trained on Indonesian-language material from media partners and public-sector websites, alongside regional-language data; the company also worked with Universitas Sumatera Utara to gather contemporary Batak-language data. The model card lists support for English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese, plus a 128,000-token context window. Its Hugging Face repository provides weights and examples for Transformers and vLLM; no hosted inference provider was listed when checked.

The model is open-weight under Google’s Gemma license, not under a conventional permissive license such as Apache 2.0 or MIT. Organizations should review the Gemma terms before adapting or redistributing it. The card estimates that running the unquantized BF16/FP16 model requires about 54–60 GB of VRAM, with setups such as one A100 80GB or multiple smaller GPUs.

GoTo’s model card reports evaluations using IndoMMLU for Indonesian-context knowledge and SEA-HELM for language capabilities. It says the evaluations were zero-shot, with 100–1,000 samples per dataset. The team also localized SEA-IFEval and SEA-MTBench; for the latter, GPT-4 served as judge against GPT-3.5 Turbo as baseline. These are publisher-reported results, not an independent replication, and the card itself cautions that outputs can hallucinate or reason inconsistently.

Busy Jakarta street with pedestrians and traffic; illustrative image, not a model screenshot.

Photo by elwis musa tambuwun on Unsplash, used under the Unsplash License.

Sources: GoTo release announcement — October 19, 2026 · GoTo-AI-Gemma-3-27B-IT model card · Gemma License · Photo by elwis musa tambuwun · Unsplash License

IndonesiaOpen-weight modelsGoTo-AIGemma 3