In brief
- PT GoTo Gojek Tokopedia has released GoTo-AI-Gemma-3-27B-IT, a 27-billion-parameter instruct model for Indonesian culture and languages, on Hugging Face.
- For Indonesian developers and public institutions, it is open-weight under Google’s Gemma licence — usable weights with licence terms to check, not unqualified “open source.”
- What’s still unknown: how it performs on your dialect and workflow; GoTo’s own benchmark suites are listed on the card, but TSN quotes no scores.
If you build apps in Bahasa Indonesia or regional languages, this is another large local model you can download and run — provided you accept the Gemma licence and can afford roughly 54–60 GB of VRAM for full-precision inference.
TechNode Global reports the model was announced on 19 October 2026 during Bali Tech Week.
What languages does it cover?
The model card lists languages supported as English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese.
It describes GoTo-AI as LLMs “pretrained and instruct-tuned for Indonesian culture and languages,” leveraging the Gemma-3 architecture. Context length is 128k tokens.
| Fact (model card) | Detail |
|---|---|
| Name | GoTo-AI-Gemma-3-27B-IT |
| Architecture | Gemma-3 |
| Initiated by | PT GoTo Gojek Tokopedia Tbk |
| Developed by | PT GoTo Gojek Tokopedia Tbk, AI Singapore |
| Licence | Gemma License |
| Context | 128k tokens |
| FP16/BF16 VRAM | about 54–60 GB |
| Suggested GPUs | 1× A100 80GB; 2× L40s; 3× RTX 4090 |
Is it open source?
Say open-weight under Google’s Gemma licence — not plain open source.
TechNode notes that GoTo “describes the release as open source, while the public model card lists Google’s Gemma license.” Organisations considering deployment should review those terms for their intended use.
Who built it?
The card says the project was initiated by GoTo and developed by GoTo and AI Singapore. Do not say GoTo built it alone.
What does GoTo say about training data?
Per TechNode (GoTo’s statements, not independently audited): training material “included Indonesian-language text from media partners and public-sector websites.” For Batak-language data, GoTo worked with Universitas Sumatera Utara.
TechNode also places the release in the broader Sahabat-AI work with Indosat Ooredoo Hutchison begun in 2024, which earlier included a 70-billion-parameter model — context for the programme, not a claim that this 27B replaces every earlier model.
What hardware do you need?
The card says FP16 or BF16 needs a minimum of approximately 54–60 GB of VRAM, recommending 1× NVIDIA A100 (80GB), 2× NVIDIA L40s, or 3× NVIDIA RTX 4090. That puts the uncompressed model beyond ordinary consumer hardware.
What should users watch for?
The card warns the model can hallucinate and that developers should do their own safety fine-tuning. TechNode similarly notes GoTo did not provide independently verified performance comparisons in the announcement it reviewed.
TSN is not quoting benchmark scores from IndoMMLU, SEA-HELM, SEA-IFEval or SEA-MTBench.
Our read
A 27B Gemma-based instruct model with Indonesian plus five regional languages is a practical sovereign-AI building block for teams that need local language coverage and can run large GPUs. The catch is licence and cost: Gemma terms are not a blank open-source cheque, and 54–60 GB VRAM keeps full-precision use in the data-centre or multi-GPU tier unless you quantise.
What’s still unknown?
- Independent third-party scorecards across the six languages
- How Batak, Balinese and other regional coverage holds up outside GoTo’s eval sets
- Production licence interpretations for commercial apps under the Gemma licence
How we sourced this
- Model facts from the Hugging Face model card for GoTo-AI-Gemma-3-27B-IT (card last modified 19 October 2026).
- Announcement framing, Bali Tech Week timing, training-data and Sahabat-AI context from TechNode Global, 19 October 2026.
- No benchmark scores quoted; “open source” wording treated as GoTo’s description, overridden for TSN by the card’s Gemma licence.
Sources
- [1] Hugging Face: GoToCompany/gemma3-27b-cpt-gotoai-v3-instruct (GoTo-AI-Gemma-3-27B-IT)
- [2] TechNode Global: GoTo releases 27B Indonesian language model, 19 October 2026
Related stories
- NVIDIA and AI4Bharat Detail a 2.8-Trillion-Token Indic Pretraining Corpus, Releasing Soon
- EmbeddingGemma 2: Google’s open multimodal embeddings for the edge
- Microsoft’s Sovereign AI Pitch: Not One Model, but the Right Model for Each Job
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
