GoTo opened a 27B model for Indonesia's regional languages. The card needs a language-by-language receipt

THE SIGNAL IN ONE SENTENCE

A model card says one system supports six languages. That sounds tidy. The people inside those six labels are not tidy. They code-switch. They use local spellings, formal registers, family expressions, technical vocabulary and jokes that move between languages halfway through a sentence. A model can look capable in one school-style benchmark and fall apart in a clinic, a village office or a customer-support chat. GoTo has released a model that puts this problem in a useful, unusually local frame. The model is called gemma3-27b-cpt-gotoai-v3-instruct. It is a 27-billion-parameter instruction-tuned system based on Google's Gemma 3 architecture. GoTo says it developed the model with AI Singapore for English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese. The weights are available through Hugging Face under the Gemma license, and the card lists a context window of 128,000 tokens. The plain signal is this: a model that names regional languages should publish a receipt for each language, each task and each speaker community. GoTo's official leaderboard already takes a step in that direction. Instead of reporting only one multilingual average, it shows columns for Indonesian, Javanese, Sundanese, Balinese and Batak. That matters because the scores are uneven. The GoTo model receives an overall Indonesian-languages average of 64.39 on the developer's composite leaderboard. Its reported language scores range from 71.09 for Javanese to 52.55 for Batak. The corresponding base Gemma 3 27B row ranges from 67.83 for Javanese to 42.64 for Batak. Those numbers suggest that continued pretraining and instruction tuning helped on the selected evaluation suite, particularly in the lowest-scoring column. They do not mean the model is 71 percent reliable in Javanese or 52 percent useful to every Batak Toba speaker. They are composite developer measurements produced from a particular collection of tasks, prompts, scoring rules and datasets. That distinction is the whole story. The leaderboard says it evaluates natural-language understanding and generation, reasoning, safety, linguistic knowledge, instruction following and multi-turn interaction. It pools evaluation data, creates 30 bootstrap samples, runs each model eight times and averages the final results. For multiple-choice work, it also normalizes scores against chance. This is better evidence than a launch claim with a single victory number. It still leaves important questions. Which dialects appear in each dataset? Who wrote or translated the prompts? What ages, regions, professions and literacy levels are represented? How often does the model preserve a user's language instead of drifting into Indonesian or English? Does it understand code-switching, or does it simply tolerate a few borrowed words? What happens when a sentence uses informal spelling, local honorifics or a script variant that was rare in training? A language name is not an evaluation plan. The model card describes two major benchmark families. IndoMMLU covers school subjects, humanities, Indonesian and local languages and cultures, social science and science, technology, engineering and mathematics. SEA-HELM covers tasks including question answering, sentiment, toxicity, translation, summarization, causal reasoning, natural-language inference and linguistic diagnostics. The card also reports instruction-following tests. Some localized datasets were translated or adapted by linguists and native speakers, which is valuable. One multi-turn evaluation uses GPT-4-1106-preview as a judge and GPT-3.5-turbo-0125 as a baseline. An AI judge can make a broad evaluation cheaper and more repeatable. It can also reward the style and preferences of another model. That is why every score should say who or what judged it, whether native speakers audited the decision and how often human and automated judgments disagreed. The useful next artifact is a language-by-language receipt. Start with the exact population represented. Do not write only Javanese. Record the regions, dialects, registers, age groups and domains in the data and evaluation. Say whether prompts came from native speakers, professional translators, public corpora, synthetic generation or a mixture. Document consent and licensing where the source permits it, and identify material that could not be traced cleanly. Then break performance down by task. A customer-service team cares about intent recognition, politeness, escalation and the faithful handling of account details. A public-health team cares about factual accuracy, uncertainty and whether a translation changes a warning. A teacher cares about curriculum alignment, age-appropriate explanations and whether local examples are treated as knowledge rather than decoration. One average cannot serve all three. The receipt should also show the mistakes, not just the score. Publish representative failures after privacy review. Count unwanted language switching, mistranslated names, erased honorifics, invented cultural claims, unsafe health advice and refusals that affect one language more than another. Let speakers report errors in their own language and make the status of each correction visible. That correction channel matters because model cards age quickly. A static table can describe the release. It cannot tell a community whether a reported problem was reproduced, fixed or ignored. Compute belongs in the receipt too. The model card estimates roughly 54 to 60 gigabytes of GPU memory for full-precision inference and suggests hardware such as one 80-gigabyte A100, two L40S cards or three RTX 4090 cards. That is a realistic warning. It is also a reminder that downloadable weights do not automatically create local access. A university lab, small language organization or regional newsroom may not own that equipment. Quantized versions can reduce the memory requirement, but compression can change quality and can affect languages unevenly. Any accessible deployment should therefore report its precision, hardware, latency, energy use and language-specific evaluation after conversion. The Gemma license matters as well. The weights are available, but that is not the same thing as software released under a conventional open-source license. Developers should read the actual terms instead of treating open access, open weights and open source as interchangeable labels. Practical uses are easy to imagine. Local businesses could route support questions without forcing everyone into English. Public-service offices could draft plain-language explanations that staff verify before publication. Teachers could generate exercises using local examples. Newsrooms could search interviews across several languages. Farmers and small merchants could ask questions in the language they already use with family and customers. Each use also carries a different harm if the model is wrong. A mistranslated product return is annoying. A mistranslated medicine dose is dangerous. A clumsy local-language joke may be embarrassing. A confident invented legal or cultural claim can travel much farther than the correction. GoTo's model card is explicit about general limitations. It warns that the system may hallucinate, produce irrelevant content and reason inconsistently. It says developers are responsible for safety fine-tuning and security in their applications. That warning should be operational, not ceremonial. High-stakes deployments need human escalation in the same language as the user. They need logs that preserve the original request and the model's answer while protecting personal information. They need tests for prompt injection, sensitive data leakage and abusive output. They need a rollback plan when an update changes behavior. They also need community governance. Native speakers should not appear only at the end as unpaid bug reporters. Language organizations, teachers, translators, creators and affected service users should help define what good performance means, which mistakes are unacceptable and which data should never be collected. The project has a strong regional premise. GoTo says the broader ecosystem is intended to support local developers and acknowledges AI Singapore along with collaborating Indonesian universities and media organizations. That is a foundation worth building on. The next release should make the partnership legible in the evidence. For each language, publish the data story, the task-level scores, the human-review protocol, the compute path, the safety failures and the correction log. Keep the official leaderboard, but add the communities behind the columns. Six language names can fit in one sentence. Six trustworthy deployments cannot. Give every language its own receipt.