News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Sahabat-AI

Sahabat-AI

Indonesian language model family by GoTo and Indosat

Sahabat-AI is a collection of large language models tuned for Indonesian and its dialects. Co-initiated by GoTo Group and Indosat Ooredoo Hutchison, it aims to support local AI development. The ecosystem includes models based on Google’s Gemma architecture, fine-tuned with data in Bahasa Indonesia, Javanese, and Sundanese.

Share on WhatsApp

  • Supports Indonesian, Javanese, and Sundanese languages
  • Based on Google Gemma2 9B architecture
  • Co-initiated by GoTo Group and Indosat Ooredoo Hutchison
  • Trained on over 448,000 Indonesian instruction pairs
  • Released under Gemma Community License

What it is

Sahabat-AI is a suite of large language models designed specifically for the Indonesian language and its regional dialects. The project is co-initiated by major Indonesian technology and telecommunications firms, GoTo Group and Indosat Ooredoo Hutchison. It includes models such as Gemma2 9B CPT Sahabat-AI v1 Instruct, which is built upon Google’s Gemma2 architecture. The system supports Bahasa Indonesia, Javanese, and Sundanese, aiming to provide culturally relevant AI tools for local developers and users.

How it came about

The Sahabat-AI ecosystem was established to empower Indonesians to build AI services using their native languages. GoTo Group and Indosat Ooredoo Hutchison launched the initiative with support from research bodies like AI Singapore and Tech Mahendra. Collaborations with leading universities, including the University of Indonesia and Bandung Institute of Technology, helped enrich the training data. Media groups such as Kompas Gramedia also contributed to ensuring local context accuracy.

Inside the model

The Gemma2 9B variant uses a decoder-only transformer design with a context window of 8192 tokens. It employs the default tokenizer from the base Gemma-2-9B model. Training involved continued pre-training followed by instruction tuning. The team used approximately 448,000 instruction-completion pairs in Indonesian, along with tens of thousands in Javanese and Sundanese. An additional 129,000 English pairs were included. Fine-tuning took about four hours on eight H100 GPUs, with two more hours for alignment.

What it can do

The model was evaluated using the SEA HELM benchmark, also known as BHASA, which tests tasks like question answering and translation. It also underwent testing on IndoMMLU for subjects ranging from humanities to STEM. English capabilities were checked against standard leaderboards. Results varied from other benchmarks due to the use of VLLM inference, which capped context at 4096 tokens during some tests. The model can hallucinate and may produce inconsistent reasoning.

How to use it

Weights are available under the Gemma Community License, which allows commercial use with certain conditions. Developers can run the model using the Hugging Face Transformers library. The project encourages contributions from researchers and developers to improve documentation and share data. Users must perform their own safety fine-tuning, as the released weights are not aligned for safety.

Safety and testing

The developers state that current Sahabat-AI models have not been aligned for safety. They advise users to implement their own security measures and fine-tuning. The team notes that the model may generate fictional elements or irrelevant content. No specific red-teaming results or safety evaluations are detailed in the provided material.

Why it matters

Sahabat-AI aims to create a local open-source ecosystem for Indonesian AI development. By providing models trained on local dialects, it seeks to lower barriers for creating applications in Bahasa Indonesia, Javanese, and Sundanese. The project invites collaboration from the global tech community to expand its capabilities and evaluation metrics.

Questions people ask about Sahabat-AI

What languages does Sahabat-AI support?

The model supports Bahasa Indonesia, Javanese, and Sundanese. It also includes training data in English to help with general language understanding and translation tasks.

Who developed Sahabat-AI?

The project is co-initiated by GoTo Group and Indosat Ooredoo Hutchison. Development involved PT GoTo Gojek Tokopedia Tbk and AI Singapore, with support from various Indonesian universities.

Is Sahabat-AI safe to use without modification?

No. The developers state that the models have not been aligned for safety. Users should perform their own safety fine-tuning and security checks before deploying the model in production.

What is the context length of the model?

The Gemma2 9B CPT Sahabat-AI v1 Instruct model has a context length of 8192 tokens. However, some evaluations used VLLM which capped this at 4096 tokens.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Sahabat-AI yet, so the facts on this page come from the sources listed above: huggingface.co/GoToCompany/gemma2-9b-cpt-sahabatai-v1-instruct. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.