# GoTo’s 27B AI Model Targets 5 Indonesian Languages [2026] > Source: > Published: 2026-10-20 16:32:13+00:00 PT GoTo Gojek Tokopedia Tbk unveiled a new 27-billion-parameter open-source language model in Sanur, Bali, on October 19, 2026, pushing deeper into a market that global AI labs have mostly treated as an afterthought: Indonesia’s regional languages. The model understands Bahasa Indonesia alongside Javanese, Sundanese, Balinese, and Bataknese, and GoTo says it will stay free for startups, businesses, universities, researchers, and public institutions to use, study, and adapt. The release matters beyond Jakarta’s tech scene. Indonesia is home to more than 700 living languages, yet almost every major [large language model](https://shattered.io/category/ai-machine-learning/) from OpenAI, Google, or Anthropic treats Bahasa Indonesia as a secondary training target and regional dialects as an afterthought at best. GoTo’s new model, part of its [Sahabat-AI open-source ecosystem](https://www.gotocompany.com/en/news/press/indosat-ooredoo-hutchison-and-goto-launch-sahabat-ai-indonesias-open-source-llm-for-empowering-digital-sovereignty), is a bet that local context and cultural nuance can be a competitive edge global labs cannot easily replicate. This is not GoTo’s first swing at the problem. The company has been building toward this moment since late 2024, and the October 19 announcement marks the most ambitious version of that project yet. Here is what GoTo actually announced, what remains unconfirmed, and why the move fits a broader pattern of countries and companies racing to build their own “sovereign AI” rather than renting it from Silicon Valley. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## GoTo Unveils Its 27-Billion-Parameter Model in Bali GoTo, the Jakarta-based super-app operator behind Gojek and Tokopedia and listed on the Indonesia Stock Exchange under the ticker GOTO, held its announcement event in Sanur, Bali, rather than its home base of Jakarta. The company’s official release describes the new system as a 27 billion parameter large language model, built to understand and generate text across Bahasa Indonesia and four regional languages: Javanese, Sundanese, Balinese, and Bataknese. Indonesian outlet [Detik](https://inet.detik.com/cyberlife/d-8716074/goto-rilis-model-llm-27b-paham-bahasa-indonesia-hingga-daerah) reported on the launch, framing it as GoTo’s answer to a gap that general-purpose chatbots have struggled to close: understanding not just Bahasa Indonesia but the linguistic texture of the country’s regional tongues. Business outlet [Bisnis.com](https://teknologi.bisnis.com/read/20261019/266/2013711/goto-luncurkan-ai-llm-open-source-27-miliar-parameter-dari-bahasa-jawa-hingga-batak) similarly covered the release, noting the model spans everything from Javanese to Bataknese and ties into GoTo’s broader open-source AI ecosystem push. GoTo has positioned the release as freely available, meaning startups, businesses, universities, researchers, and public institutions can use, study, and adapt the model without the licensing friction that often comes with proprietary systems from Western AI labs. That framing echoes the open-weight strategy used by companies ranging from Mistral to DeepSeek, where releasing model weights publicly builds developer goodwill and distribution even without a direct revenue line from the model itself. Financial outlet [Kontan](https://industri.kontan.co.id/news/goto-luncurkan-ekosistem-ai-lewat-model-27-miliar-parameter) reported that the 27-billion-parameter model carries the Hugging Face identifier `gemma3-27b-cpt-gotoai-v3-instruct`, a naming convention that suggests GoTo built the system on top of an existing open-weight foundation model rather than training a fully custom architecture from the ground up. GoTo has not detailed the exact training data volume, hardware stack, or licensing terms attached to the release, and those specifics remain unconfirmed as of this writing. ``` Model identifier (per Kontan): gemma3-27b-cpt-gotoai-v3-instruct Parameters: 27 billion Languages: Bahasa Indonesia, Javanese, Sundanese, Balinese, Bataknese Release date: October 19, 2026 Location: Sanur, Bali Access: Free for startups, businesses, universities, researchers, public institutions ``` ## Inside the Numbers: What a 27B Model Can Realistically Do Parameter count is a blunt instrument for judging a model’s capability, but it still offers a useful frame of reference. A 27-billion-parameter model sits in the mid-tier of today’s open-weight landscape, big enough to handle nuanced language tasks like translation, summarization, and conversational assistance, but small enough to run on a single high-end GPU rather than requiring a data center’s worth of accelerators. That tradeoff matters for a country like Indonesia, where cloud infrastructure costs and connectivity vary sharply between Jakarta’s data centers and more remote provinces. The model’s Gemma-3 lineage, as implied by its Hugging Face name, points to Google’s open-weight Gemma family as the likely base architecture GoTo fine-tuned with Indonesian and regional-language data. This approach, known as continued pre-training (the “cpt” in the model identifier), lets a company take a strong general-purpose foundation and adapt it to a specific linguistic or cultural domain without the enormous compute budget required to train a frontier model from scratch. It is the same playbook used by sovereign-AI projects elsewhere, including Saudi Arabia’s [Humain building a 428-billion-parameter Arabic LLM on MiniMax’s architecture](https://shattered.io/saudi-humain-428b-arabic-llm-minimax-2026/). What GoTo has not disclosed is just as telling as what it has. The company has not released benchmark scores against standard evaluation suites, has not specified a context window length, and has not named the engineers or researchers who led the project. Until GoTo or independent researchers publish benchmark results, direct performance comparisons with global players remain speculative at best. ## From 2024’s Sahabat-AI to the 2026 Expansion The October 19 release did not appear out of nowhere. GoTo and telecom operator Indosat Ooredoo Hutchison first introduced the Sahabat-AI ecosystem on November 14, 2024, pitching it as Indonesia’s own open-source LLM initiative built to serve Bahasa Indonesia and the country’s regional languages. That original release made its models freely available through Hugging Face, setting the open-access tone that the 2026 model continues. GoTo’s own description of the initiative has stayed consistent across both releases: the appeal of Sahabat-AI lies in deep localization, built specifically for Indonesian and regional languages like Javanese and Sundanese, rather than treating them as a translation layer bolted onto an English-first model. The gap between the 2024 launch and the 2026 model shows how quickly the underlying technology has scaled, even as the company’s language-coverage goals have stayed remarkably stable. | Milestone | Date | Scale | Languages Covered | |---|---|---|---| | Sahabat-AI ecosystem launch | November 14, 2024 | 8B and 9B parameter models | Bahasa Indonesia, regional languages | | Multilingual chat expansion | Reported mid-2025 | 70 billion parameters | Bahasa Indonesia, Javanese, Sundanese, Balinese, Bataknese, plus international languages | | GoPay integration | 2025 | N/A (product integration) | Conversational assistance inside GoTo’s payments app | | New open-source LLM release | October 19, 2026 | 27 billion parameters | Bahasa Indonesia, Javanese, Sundanese, Balinese, Bataknese | The jump from a 70-billion-parameter model reported in 2025 down to a 27-billion-parameter release in 2026 looks like a step backward on paper, but it likely reflects a different design goal rather than a retreat. Smaller open models are cheaper to run, easier for third-party developers to fine-tune on modest hardware, and faster to deploy inside products like GoTo’s own ride-hailing and e-commerce apps. GoTo has not confirmed whether the 27B model replaces the earlier 70B system or runs alongside it for different use cases. ## Why Javanese, Sundanese, Balinese, and Bataknese Matter Indonesia’s linguistic map is one of the most complex on Earth. Javanese alone has roughly 68 million native speakers, more than most European languages, yet it barely registers in the training data of mainstream global AI models. Sundanese, Balinese, and Bataknese each carry millions of speakers and centuries of literary and oral tradition that general-purpose chatbots routinely flatten into generic Bahasa Indonesia or, worse, broken English transliteration. That gap creates a real product problem for companies trying to reach Indonesian users outside the major cities. A customer service bot that only speaks formal Bahasa Indonesia will frustrate a user in rural Central Java who expects to communicate in Javanese, or a Balinese tourism operator who needs region-specific cultural references to sound authentic. GoTo’s bet is that solving this problem directly, rather than hoping a bigger global model eventually gets around to it, gives the company a defensible advantage in its home market. This logic lines up with a point GoTo made when it first introduced Sahabat-AI in 2024: deep localization for a language and its regional variants is the kind of work a company headquartered in Jakarta can do faster and more credibly than a lab in San Francisco or Mountain View. Whether that logic scales cleanly from an 8-billion-parameter prototype to a 27-billion-parameter production model is one of the open questions the October 19 release leaves unanswered. ## Open Source as Strategy, Not Charity GoTo is giving this model away, and that choice is strategic rather than altruistic. Open-weight releases let a company build developer mindshare, attract academic citations, and seed an ecosystem of downstream applications it does not have to build or fund itself. Universities and research centers can fine-tune the model for specialized tasks. Startups can bolt it into customer-facing products without paying API fees to OpenAI or Anthropic. Government agencies can run it on local infrastructure rather than routing citizen data through a foreign cloud provider. That last point lines up with language GoTo used when it launched the original Sahabat-AI ecosystem in 2024. The company’s leadership framed the project as a way to put the power of AI into the hands of everyone in Indonesia, and argued that operating natively in Bahasa Indonesia closes context and cultural reference gaps that global large language models leave open. GoTo also called on businesses, government ministries, universities, and other institutions to help develop the ecosystem rather than treating it as a closed, GoTo-only product. Giving away a 27-billion-parameter model is not free for GoTo to produce, even if it is free for everyone else to use. Training and fine-tuning runs at that scale require meaningful GPU budgets, and GoTo has not disclosed what it spent building the 2026 version. The payoff, if the open-source bet works, is less about direct licensing revenue and more about cementing GoTo as the default AI infrastructure layer for Indonesian-language applications, the same way companies elsewhere have tried to capture [regional AI leadership through open releases rather than paywalled APIs](https://shattered.io/jais-3-arabic-ai-models-375b-params-2026/). ## GoTo’s Business Case: From Ride-Hailing to AI Infrastructure GoTo’s core businesses, Gojek’s ride-hailing and delivery network and Tokopedia’s e-commerce marketplace, generate enormous volumes of Indonesian-language text and voice data every day: customer support chats, driver-passenger messages, product reviews, and seller communications. That data, handled responsibly, is exactly the kind of raw material needed to train a model that actually understands how Indonesians write and speak across dialects, rather than relying on formal, textbook Bahasa Indonesia scraped from government documents and news articles. Embedding Sahabat-AI into GoTo’s own products is a logical next step, following the pattern set after the 2024 and 2025 releases when the earlier Sahabat-AI model was integrated into the GoPay payments app for conversational assistance. A more capable, more efficient 27-billion-parameter model could power customer support automation across Gojek and Tokopedia, reducing the operating costs that come with staffing human support agents at the scale GoTo operates. There is also a defensive angle. As global AI assistants from Google, OpenAI, and Chinese labs expand into Southeast Asia, GoTo risks becoming just another app layered on top of someone else’s AI infrastructure. Owning a credible, Indonesian-first model gives GoTo a hedge against that dependency, similar to the logic behind [the broader open-source AI bet that companies like Mistral and Reflection have made](https://shattered.io/ai-open-source-bet-86-percent-beam-mistral-2026/) to avoid being locked into a handful of US-based frontier labs. ## Indonesia Joins a Global Sovereign-AI Race GoTo’s release lands squarely inside a pattern playing out across emerging markets in 2026: countries and regional champions building their own language-specific AI models rather than depending entirely on US or Chinese frontier labs. Saudi Arabia’s Humain built a 428-billion-parameter Arabic model on MiniMax’s architecture. The Jais project has pushed out a family of Arabic-language models spanning 375 billion parameters across six variants. India’s IIT Madras has shipped open Bharat-focused models through its Bodhan AI initiative. Each of these projects shares GoTo’s underlying thesis: a model trained with deep attention to a specific language and culture can outperform a general-purpose global model on local tasks, even without matching it on raw parameter count. | Project | Country/Region | Reported Scale | Primary Language Focus | |---|---|---|---| | GoTo Sahabat-AI (2026 release) | Indonesia | 27 billion parameters | Bahasa Indonesia, Javanese, Sundanese, Balinese, Bataknese | | Humain | Saudi Arabia | 428 billion parameters | Arabic | | Jais 3 | UAE / Gulf region | 375 billion parameters across 6 models | Arabic | | Bodhan AI | India | 4 open models | Indian regional languages | The common thread across these efforts is a rejection of the assumption that bigger, English-centric models will eventually “trickle down” adequate support for every regional language. Instead, each country is choosing to build its own localized layer now, rather than waiting for Silicon Valley or Beijing to prioritize a few hundred million speakers of a language that is not English, Mandarin, or Spanish. ## Industry Voices on Sahabat-AI’s Mission When GoTo and Indosat Ooredoo Hutchison first launched the Sahabat-AI ecosystem in November 2024, GoTo chief executive Patrick Walujo laid out the thinking that still underpins the 2026 expansion. Our vision for Sahabat-AI is to put the power of AI into the hands of everyone in Indonesia, Walujo said at the time, according to GoTo’s official launch announcement. Walujo went further in comments reported by Marketing Interactive, arguing that by operating in Bahasa Indonesia, Sahabat-AI addresses critical context and cultural reference gaps left by the global large language models. He described the practical stakes in concrete terms, saying the project would help businesses communicate in new ways with customers, help government ministries develop tools to engage with citizens more comprehensively, and drive meaningful change for millions of people across the country. Walujo also framed the ecosystem as an open invitation rather than a closed GoTo product, telling stakeholders from all sectors to join the development of the open-source LLM ecosystem as the country aims to move forward together. NVIDIA, which has supported sovereign-AI projects across Southeast Asia, has described Sahabat-AI as a collection of open-source Indonesian large language models that local industries, government agencies, universities, and research centers can use to create generative AI applications, a characterization published on the company’s blog covering Indonesian AI initiatives. That framing captures the core pitch GoTo is making again with the 27-billion-parameter release: infrastructure for an entire national AI ecosystem, not just a product feature for GoTo’s own apps. ## What GoTo Has Not Confirmed Several details that would normally accompany a major model release are missing from GoTo’s public statements so far. The company has not published benchmark scores comparing the 27-billion-parameter model against other open-weight systems in its class, such as Meta’s Llama family, Alibaba’s Qwen line, or Mistral’s smaller models. It has not disclosed the volume or sourcing of training data used for the regional-language fine-tuning, an important detail given how sensitive language data collection can be in culturally and religiously diverse communities. It has not named a specific license (whether Apache 2.0, a custom open license, or something more restrictive), and it has not identified the engineering or research team responsible for the project. The company also has not clarified whether the 27-billion-parameter release supersedes the roughly 70-billion-parameter Sahabat-AI model reported in 2025, or whether both versions will continue to exist for different use cases. Reports have not specified a release cadence for future updates, nor have they confirmed what compute infrastructure GoTo used for training and fine-tuning. Readers should treat any claim about the model’s relative performance against global systems as unverified until GoTo or an independent benchmark publishes comparative results. ## Market and Investor Reaction GOTO shares trade on the Indonesia Stock Exchange, and the company’s AI strategy sits alongside its core ride-hailing, delivery, and e-commerce businesses in investor evaluations. GoTo has spent the past two years working to show profitability across its core operations after years of growth-at-any-cost spending, and AI investment represents a parallel bet: that owning language-specific AI infrastructure strengthens the company’s long-term competitive position even if it does not generate direct revenue in the near term. Regional telecom and technology investors have watched sovereign-AI announcements closely in 2026, partly because governments across Southeast Asia have signaled interest in data localization and reduced reliance on foreign AI providers for sensitive government and financial applications. A credible, homegrown open-source model gives GoTo a stronger pitch when selling AI infrastructure services to Indonesian government agencies or regional banks that face regulatory pressure to keep AI processing within national borders, a dynamic playing out in comparable form around [the pricing and access gaps that have opened between premium global AI assistants](https://shattered.io/claude-vs-chatgpt-vs-gemini-price-gap-2026/) and more accessible regional alternatives. ## How Sahabat-AI Stacks Up Against Global Open Models Direct performance comparisons remain speculative without published benchmarks, but the competitive landscape GoTo is stepping into is crowded. Mistral’s Le Chat family, DeepSeek’s V4.1 line, and Alibaba’s Qwen models all offer multilingual support that covers Bahasa Indonesia to varying degrees, though none specialize in the depth of regional dialect coverage GoTo is targeting with Javanese, Sundanese, Balinese, and Bataknese. That specialization is both the new model’s biggest potential strength and its biggest unproven claim. Open-weight releases from companies like [Reflection, which opened its Beam model under an Apache 2.0 license with more than 40 partners](https://shattered.io/reflection-beam-open-source-apache-2-0-40-partners-2026/), show that giving away a capable model can still build a commercial ecosystem around services, fine-tuning, and enterprise deployment support. GoTo appears to be following a similar playbook: the model itself is free, but the infrastructure, integration services, and enterprise support built around it could become a genuine revenue line as more Indonesian institutions adopt the technology. The parameter count gap between GoTo’s 27-billion model and projects like Humain’s 428-billion-parameter Arabic system also raises a fair question about ambition versus efficiency. A smaller model that runs cheaply and deploys widely across Indonesia’s uneven infrastructure may ultimately reach more real users than a larger model that only a handful of well-funded institutions can afford to run. ## Historical Context: Indonesia’s Broader AI Push Indonesia’s government has pushed digital sovereignty as a policy priority for several years, citing concerns about data residency, cybersecurity, and economic dependence on foreign technology platforms. GoTo’s AI strategy fits inside that broader national conversation, positioning the company as a private-sector champion for homegrown AI infrastructure at a time when policymakers across the region are weighing rules on where AI training data can be stored and processed. The 2024 partnership between GoTo and Indosat Ooredoo Hutchison already signaled this ambition, pairing GoTo’s consumer data and product reach with Indosat’s telecom infrastructure and government relationships. The 2026 release suggests that partnership, or at least its underlying strategy, has continued to mature, even as GoTo appears to be driving this specific release independently based on available reporting. ## Five Predictions for Sahabat-AI’s Next Phase - GoTo will likely integrate the 27-billion-parameter model more deeply into Gojek and Tokopedia customer support within the next two to three quarters, following the same pattern it used when the earlier Sahabat-AI model reached the GoPay app. - Expect independent developers and Indonesian universities to publish the first third-party benchmark comparisons within weeks of the Hugging Face release, since the model’s free availability invites exactly that kind of community testing. - GoTo will face pressure to clarify licensing terms and training data sourcing as adoption grows, particularly if government agencies consider deploying the model for citizen-facing services. - Other Southeast Asian nations, including the Philippines, Vietnam, and Malaysia, may accelerate their own regional-language AI projects in response, following the same competitive logic driving the Gulf region’s Arabic-language model race. - GoTo is likely to position future Sahabat-AI releases as a recurring platform rather than a one-off event, mirroring how Mistral, DeepSeek, and Reflection have each shipped iterative open-weight updates roughly every few months in 2026. ## What This Means for Southeast Asia’s AI Market Southeast Asia has largely been a consumption market for AI technology built elsewhere, with companies and governments licensing or deploying models trained primarily on English and Mandarin-language data. GoTo’s release, following its 2024 and 2025 predecessors, signals a shift toward local production of foundational AI infrastructure rather than pure consumption. If the strategy works, it could pressure other regional tech giants, from Singapore’s Grab to Vietnam’s VNG, to pursue similar open-source language investments rather than ceding that ground entirely to global labs. It also raises the stakes for how global AI labs approach Southeast Asian languages going forward. If GoTo’s regional-dialect model proves genuinely useful for government, education, and commercial applications, it becomes harder for companies like OpenAI, Google, and Anthropic to treat Bahasa Indonesia and its regional variants as a low-priority afterthought in their own training pipelines. Competition, even from a single mid-sized [open model](https://shattered.io/import-ai-476-jack-clark-china-open-weight-models-2026/), tends to accelerate investment across an entire category. ## Frequently Asked Questions ### What did GoTo actually announce on October 19, 2026? GoTo announced a new 27-billion-parameter open-source large language model at an event in Sanur, Bali. The model understands Bahasa Indonesia along with four regional languages: Javanese, Sundanese, Balinese, and Bataknese, and GoTo says it will remain freely available for startups, businesses, universities, researchers, and public institutions. ### Is this the same thing as Sahabat-AI? Yes. The 27-billion-parameter model is the latest release within GoTo’s Sahabat-AI open-source ecosystem, which GoTo and Indosat Ooredoo Hutchison first launched in November 2024 with smaller 8-billion- and 9-billion-parameter models. ### Where can I download or access the model? Reporting from Kontan indicates the model is listed on Hugging Face under the identifier gemma3-27b-cpt-gotoai-v3-instruct. GoTo has described the release as freely available for use, study, and adaptation by eligible organizations and individuals. ### Is the 27-billion-parameter model bigger or smaller than earlier Sahabat-AI versions? It is smaller than the roughly 70-billion-parameter Sahabat-AI model reported in 2025, but considerably larger than the original 8-billion- and 9-billion-parameter models from the 2024 launch. GoTo has not confirmed whether the new model replaces the 70-billion-parameter version or coexists with it for different applications. ### Why does GoTo care about regional languages like Javanese and Bataknese? Indonesia’s regional languages collectively represent tens of millions of speakers who global AI models typically underserve. GoTo has argued that deep localization for Bahasa Indonesia and its regional variants closes cultural and contextual gaps left by general-purpose global language models, giving the company a product advantage with Indonesian users outside major urban centers. ### Does GoTo plan to use this model inside Gojek or Tokopedia? GoTo has not confirmed specific product integration plans for the October 2026 release, but an earlier version of Sahabat-AI was integrated into the GoPay payments app for conversational assistance, suggesting similar integration into GoTo’s other products is plausible. ### How does this compare to other sovereign AI projects globally? GoTo’s release follows a pattern seen in Saudi Arabia’s 428-billion-parameter Humain model, the 375-billion-parameter Jais 3 Arabic model family, and India’s Bodhan AI initiative. Each project focuses on deep localization for a specific language and culture rather than competing on raw scale with global frontier labs. ### Is the model open source or just free to use? GoTo has described the release as freely available for startups, businesses, universities, researchers, and public institutions to use, study, and adapt. The company has not confirmed the specific open-source license governing the model, so the exact terms of redistribution and commercial use remain unconfirmed.