Introducing Sahabat AI API on Neosantara (Experimental)

Neosantara is launching experimental API support for the Sahabat AI model ecosystem — the first Indonesian LLM family, continued-pretrained on 50 billion Indonesian tokens and embedded into GoTo's Dira AI voice assistant, now serving Indonesia's largest digital economy platform (GoTo press release, 2024). This initiative brings two high-performance Indonesian models to developers through a single OpenAI-compatible endpoint, with transparent Rupiah billing and local gateway latency.

Indonesia's AI market reached an estimated USD 1.1B in 2026, growing at 38% CAGR, with the LLM segment alone projected to hit USD 38.7B by 2031 (IDC 2025; Mobility Foresights, 2025). Localized models are no longer optional — they are becoming the default for Indonesian-language AI applications.

Unlike global language models that require extensive prompt engineering to produce culturally aware Indonesian outputs, Sahabat AI models are continued-pretrained on 50 billion Indonesian text tokens. This means they capture formal, informal, and regional language patterns out of the box — no custom system prompts needed for local context. In independent evaluations comparing Sahabat AI against SEA-LIONv3 across 50 Indonesian-language challenges, Sahabat AI took a clear lead in linguistic fluency and domain-specific accuracy (SUPA AI Blog, 2025).

#Why Localized Sahabat AI Models Outperform General LLMs for Indonesian Tasks

The performance gap between general-purpose and localized models is measurable. On the SEA-HELM benchmark — a standardized evaluation across 7 Southeast Asian languages — Sahabat AI v1-9B Instruct scores 64.15 for Bahasa Indonesia, 64.44 for Javanese, and 54.91 for Sundanese, outperforming baseline models on culturally grounded tasks (SEA-HELM leaderboard, 2025).

On the MMLU-ProX benchmark, Indonesian-language reasoning scored 81.3% with top multilingual models, placing Indonesia ahead of all other East and Southeast Asian languages tested — but only when using models with significant Indonesian training data (EMNLP 2025).

General models like GPT-4o and Claude perform adequately on conversational Indonesian, but their performance degrades sharply on formal Indonesian legal and regulatory text — the kind developers need for production systems (AIMenta, 2026). Sahabat AI models close this gap because they are trained on a corpus that mirrors Indonesia's linguistic diversity: 700+ regional languages, Javanese (95M speakers), Sundanese (42M), and formal-register Indonesian that differs substantially from social-media training data.

The LORAXBENCH study (EMNLP 2025) evaluated multilingual models across 20 Indonesian languages and six tasks. Sahabat AI, alongside Gemini 1.5 Pro and Gemma 9B, showed the strongest cross-language performance — including on low-resource languages like Buginese and on formal politeness registers like Javanese Krama, which most general models handle poorly.

These models also help organizations comply with Indonesia's PDP Law, which took effect in October 2024 with enforcement expected by October 2026 — routing traffic through local gateways prevents data egress across international borders, a key compliance requirement (AIMenta, 2026).

#Which Sahabat AI Model Should You Use?

Both models are available through Neosantara's experimental API, each optimized for different workloads:

The Lion 9B variant (a community-merged model that ranked #3 on the Open LLM Leaderboard for sub-10B models in November 2024) excels in open-ended conversation — independent tests show it outperforms SEA-LIONv3 by a significant margin in recognizing regional slang and interpreting idiomatic expressions (SUPA AI Blog, 2025). The Llama 3 CPT variant (official GoToCompany instruct model, continued-pretrained from AI Singapore's Sea-Lion v2.1 on 50B tokens) is better suited for structured tasks where precise instruction following matters. Both support an 8,192 token context window, sufficient for summarizing Indonesian documents or processing multi-turn conversations.

If you are already using Neosantara for other models, these Sahabat AI variants work through the same API endpoint — just change the model parameter.

#How Integration Works: One API, Two Models

Neosantara's OpenAI-compatible API means you can integrate Sahabat AI models with your existing codebase in minutes. No separate SDK, no special configuration — just a model ID swap.

import openai

client = openai.OpenAI(
    api_key="your_neosantara_api_key",
    base_url="https://api.neosantara.xyz/v1"
)

response = client.chat.completions.create(
    model="sahabatai-lion-9b-ties-v1",
    messages=[
        {"role": "system", "content": "You are a helpful assistant fluent in Indonesian and local cultural context."},
        {"role": "user", "content": "Write a short summary of the history of Jakarta."}
    ],
    temperature=0.7
)

print(response.choices[0].message.content)

Compatibility note: Because Neosantara uses the OpenAI protocol, Sahabat AI models work with any tool that supports OpenAI-formatted requests — including LiteLLM, any-llm, and custom HTTP clients. If you already use LiteLLM with Neosantara or any-llm with Neosantara, adding Sahabat AI requires only a model parameter change.

Because this integration is experimental, model latency and server availability may change during the test phase. Neosantara recommends setting up fallback configurations in your production code — for example, routing to a stable model like Kimi K2 or Agno-based agents when the Sahabat AI endpoint is unavailable.

#Indonesia LLM Landscape: Where Sahabat AI Fits

The Indonesian LLM ecosystem has matured rapidly. For context, the Indonesia NLP market was valued at USD 790.2M in 2024 and is projected to reach USD 9.4B by 2033 (30.5% CAGR) (DataCube Research, 2025). Three tiers of models are available:

Sahabat AI occupies a unique slot: it is the only production-grade Indonesian LLM embedded into a major consumer platform. GoTo's Dira voice assistant — first launched on Google Gemini 1.5 Flash in July 2024, then transitioned to Sahabat AI infrastructure by November 2024 — handles Gojek and GoPay interactions in native languages and dialects (NVIDIA blog, 2024). In June 2025, GoTo and Indosat launched a 70-billion parameter Sahabat AI model supporting 5 local languages (Indonesian, Javanese, Sundanese, Balinese, Bataknese), deepening the ecosystem's reach (GoTo press release, 2025). For Indonesian businesses, the choice is not between "better model" and "local model" — it is about matching the model to the use case. General-purpose frontier models remain the best choice for complex reasoning; Sahabat AI is the better choice when linguistic and cultural accuracy matters.

#Frequently Asked Questions

#What does "experimental" mean for production use?

Experimental status means the integration is in active development. Performance, latency, and model availability may change during the testing phase. Neosantara recommends implementing model fallbacks in your production architecture — for instance, routing to a stable general-purpose model if the Sahabat AI endpoint returns errors.

#Are these models free to use with the Neosantara free tier?

Yes. Sahabat AI models are available through the same Neosantara API key and free credit system. Sign up to receive Rp 10,000 in free credit to test both models against your use case before committing to a paid plan.

#Can I use Sahabat AI models through LiteLLM or any-llm?

Yes. Neosantara's OpenAI-compatible API endpoint means Sahabat AI works with any tool that supports OpenAI-formatted requests — including LiteLLM, any-llm, LangChain, and custom clients. The only change is the model parameter.

#What's the difference between Lion 9B and Llama 3 CPT?

Lion 9B prioritizes natural conversational fluency with cultural awareness, making it ideal for chatbots and interactive assistants. It outperforms regional alternatives on informal language and slang recognition. Llama 3 CPT is optimized for structured instruction following — better suited for classification, extraction, and structured data tasks where precise output formatting is required.

#How does Sahabat AI compare to SEA-LION?

In an independent 50-challenge evaluation across language, domain knowledge, geography, and combined tasks, Sahabat AI outperformed SEA-LIONv3 in linguistic fluency and domain-specific accuracy, while SEA-LIONv3 held a slight edge in geographical queries (SUPA AI Blog, 2025). Both models struggle with tasks requiring integration of multiple competencies, indicating room for improvement across the board.

#Source References

1. GoTo Company, "GoTo and Indosat Ooredoo Hutchison Launch Sahabat-AI," November 2024. https://www.gotocompany.com/en/news/press/goto-launches-sahabat-ai-enhancing-its-leadership-in-indonesias-technology-sector
2. GoTo Company, "Sahabat-AI 70B Model Launch with Multilingual Chat Service," June 2025. https://www.gotocompany.com/en/news/press/sahabat-ai-gets-smarter-indosat-and-goto-launch-new-70-billion-parameter-model-with-multilingual-chat-service
3. NVIDIA Blog, "Indonesia Tech Leaders Build Sovereign AI with Sahabat-AI," November 2024. https://blogs.nvidia.com/blog/indonesia-tech-leaders-sovereign-ai/
4. Neosantara, "Models Overview," 2026. https://docs.neosantara.xyz/en/models-overview
5. Neosantara, "Quickstart Guide," 2026. https://docs.neosantara.xyz/en/quickstart
6. SUPA AI, "Benchmarking Bahasa Indonesia LLMs: SEA-LIONv3 vs SahabatAI-v1," February 2025. https://blog.supa.so/2025/02/21/benchmarking-bahasa-indonesia-llms-sea-lionv3-vs-sahabatai-v1/
7. EMNLP 2025, "MMLU-ProX: A Multilingual Benchmark for Advanced LLM Evaluation," 2025. https://aclanthology.org/2025.emnlp-main.79.pdf
8. EMNLP 2025, "LORAXBENCH: A Multitask, Multilingual Benchmark Suite for 20 Indonesian Languages," 2025. https://aclanthology.org/2025.emnlp-main.881.pdf
9. SEA-HELM Leaderboard, "Southeast Asian Languages Evaluation," 2025. https://leaderboard.sea-lion.ai/detailed/ID
10. AIMenta, "Indonesia Enterprise AI 2026 — PDP Law, Digital Economy, Bahasa Indonesia," April 2026. https://aimenta.ai/insights/indonesia-enterprise-ai-2026-digital-economy-regulation
11. DataCube Research, "Indonesia AI Natural Language Processing Market Size," November 2025. https://www.datacuberesearch.com/indonesia-natural-language-processing-market

All URLs retrieved June 2026.