The Myanmar AI Gap

1Where Burmese stands

About 43 million people speak Burmese.^⁠[1] The language is no longer missing from AI. Google Translate added it in December 2014.^⁠[2] Google lists Burmese among the languages every Gemini model can read and answer in,^⁠[3] ChatGPT offers a Burmese interface,^⁠[4] and Meta's NLLB translation model covers it.^⁠[5] On BURMESE-SAN, a Burmese benchmark published in 2026, Gemini 2.5 Pro scored 72.35 percent and GPT-5 scored 66.46 percent, up from 51.61 percent for GPT-4o.^⁠[1]

Support is thinner than those headlines suggest. The Gemini app's own list of supported languages leaves out Burmese, while it includes Khmer and Lao.^⁠[6] Meta's Llama 4 officially supports Thai, Vietnamese, Indonesian and Tagalog, and not Burmese.^⁠[7] Open models are the ones local developers can download, adapt and run on their own hardware, so their gaps matter most for the products a country builds for itself.

So the useful question is not whether AI can handle Burmese. It is how well it works compared with other languages, who builds it, and whether anyone in Myanmar can shape it. On all three, the evidence points the same way. Burmese has a sliver of the text that models learn from. It scores far below its neighbours on the same models. And Myanmar's capacity to build its own tools has shrunk while its neighbours built theirs.

2Too little text

Language models learn from text, most of it collected from the public web. Common Crawl, the open archive behind most public training sets, labels each page it saves by language. In its latest crawl, from August 2026, Burmese is 0.0164 percent of pages. Thai is 0.374 percent, about 23 times as much. Vietnamese is 1.075 percent, about 66 times as much, and English is 40.45 percent.^⁠[8]

The training sets built from that archive inherit the gap. Google's mC4 has 0.9 billion Burmese tokens, against 11 billion for Thai and 2,733 billion for English.^⁠[9] FineWeb2, released in 2025, has 12.35 GB of Burmese text and 278.68 GB of Thai.^⁠[10] When AI Singapore trained its first SEA-LION model for the region's languages, Burmese made up 0.49 percent of the training tokens, and reaching even that meant repeating the same 1.2 billion tokens four times.^⁠[11]

Wikipedia shows how few people are writing. In September 2026 the Burmese Wikipedia had 286 active users. The Thai Wikipedia had 2,727 and the Vietnamese 4,955.^⁠[12]

Burmese also carries a problem of its own. For most of the last decade, most Burmese text was typed in Zawgyi, a font encoding that uses the Unicode code points for Myanmar script in its own, incompatible way. Text written in one looks garbled in the other.^⁠[13] Myanmar officially moved to Unicode on 1 October 2019, when an estimated 90 percent of people still used Zawgyi.^⁠[14] Much of the Burmese text on the web was written before the switch. The team behind Google's MADLAD-400 dataset found that "a large fraction of Myanmar script data on the internet is Zawgyi encoded data" and had to convert it,^⁠[15] and in the CC-100 corpus the Zawgyi file for Burmese is nearly four times the size of the Unicode one.^⁠[16] A dataset that skips this step can lose that text, or learn from scrambled characters.

Researchers have a name for languages in this position. In a 2020 study that sorted the world's languages by the resources available for them, Joshi and colleagues placed Burmese in class 1 on a scale from 0 to 5, the "Scraping-Bys": languages with some unlabelled text and almost no labelled data. Thai is in class 3 and Vietnamese in class 4. The study singled out Burmese as a language with millions of speakers and almost no research attention.^⁠[17]

3The gap inside the models

The clearest test gives one model the same questions in different languages. Belebele does this with 900 reading comprehension questions, each with four possible answers, in 122 language variants, so random guessing scores 25.^⁠[18] In the Belebele paper, Llama 2 70B scored 90.9 in English, 46.2 in Thai and 24.1 in Burmese. GPT-3.5 Turbo, answering without examples, scored 87.7, 55.7 and 30.3.^⁠[18]

Those are 2023 models, and the best systems have moved on. BURMESE-SAN measured the change: GPT-4o scored 51.61 percent, GPT-5 66.46 percent, and Gemini 2.5 Pro led at 72.35 percent. Open models trail. The best open model in the study, ERNIE 4.5, scored 54.68 percent, and Llama 3.3 70B scored 23.07 percent.^⁠[1]

The size of the gap now depends on who built the model. On AI Singapore's SEA-HELM leaderboard, in its August 2026 version, Google's Gemma 4 31B scores 69.13 in Burmese and 71.09 in Thai. Qwen 3.6 27B scores 47.92 in Burmese and 68.85 in Thai, and NVIDIA's Nemotron 3 Super scores 6.17 in Burmese and 59.28 in Thai.^⁠[19] The tasks differ between languages, so these are not exact comparisons, but the pattern is plain: a few builders treated Burmese as a priority, and others did not.

Burmese also costs more to run. Models read text as tokens, and a tokenizer built mostly from English breaks Burmese into many small pieces. Petrov and colleagues found that the tokenizer used by GPT-4 needs 11.7 times as many tokens for a Burmese sentence as for the same sentence in English, one of the highest ratios of any language they measured. Tokenizers designed for many languages, such as mT5's, need about 1.6 times.^⁠[20] Every extra token costs money and fills the model's limited context, so the same number of tokens carries "less than a tenth of the content" in Burmese.^⁠[20] The fix exists. It depends on whoever designs the tokenizer choosing to use it.

4Why the text does not grow

More Burmese text would exist if more people in Myanmar could get online, stay online and publish. Since 2021, each of those has become harder.

Myanmar recorded more internet shutdowns than any other country in 2024, with 85,^⁠[21] and again in 2025, with 95. Since 2021, shutdowns have reached all 14 of its states and regions.^⁠[22] Freedom House reports that mobile networks must block every website not on an approved list, and that a block on VPNs in 2024 also disrupted services from Google and Amazon.^⁠[23] Text that is never typed never reaches a training set.

Even counting who is online has become difficult. The official internet use figures that the World Bank publishes from the International Telecommunication Union stop at 45 percent for Myanmar in 2020, with nothing since. The same series puts Thailand at 91 percent and Viet Nam at 84 percent in 2024.^⁠[24] The World Bank's own estimate is that 44 percent of Myanmar's people were online in January 2025.^⁠[25]

Power is the next limit. Myanmar generated an average of 3,034 MW between January and September 2025, against an installed capacity of 6,880 MW.^⁠[25] In June 2026 the World Bank described generation of roughly 3,000 MW a day, covering only half of demand, and firms surveyed in October 2025 reported a median outage of four hours.^⁠[26] Training and serving models needs steady power, and so does every office, school and phone that would produce Burmese text.

The economy that would pay for all of this has shrunk. Real GDP fell 9.0 percent in fiscal year 2020/21 and 12.0 percent in 2021/22.^⁠[25] The World Bank expects it to stay about 11 percent below its level before the pandemic,^⁠[27] and estimates that 29.9 percent of people lived in poverty in 2025.^⁠[26] Myanmar has been on the Financial Action Task Force's list of high risk jurisdictions since October 2022,^⁠[28] a listing the World Bank says raises the cost of trade finance and capital flows, and the authorities limit payments abroad that are not for trade.^⁠[25,26]

5The people who would build it

AI is built by people, and the supply of people has been cut at every stage. Myanmar's schools were fully closed for 532 days between February 2020 and February 2022, the longest closure in East Asia and the Pacific. In 2021, about 30 percent of teachers were dismissed. The share of young people aged 6 to 22 in education fell from 69.2 percent in 2017 to 56.8 percent in 2023.^⁠[29] ISP-Myanmar, a research group, reports that candidates for the matriculation exam fell from over 900,000 in 2020 to roughly 250,000 in 2026.^⁠[30]

Research output shows the same break. Myanmar published 274 scientific and technical journal articles in 2019 and 234 in 2023. Over the same years Viet Nam went from 5,684 to 10,644, Thailand from 12,304 to 16,656, and Cambodia from 137 to 201.^⁠[31] In 2019 Viet Nam published about 21 articles for every one from Myanmar. By 2023 it published about 45.

The people trained before 2021 are leaving. In a World Bank survey of 2,400 employed university graduates in early 2024, 35 percent said they would take a similar job abroad for the same pay.^⁠[32] The World Bank's firm survey found resignations linked to migration rising to 16 percent in March 2026, and 22 percent in Yangon, with 53 percent of those who resigned going abroad.^⁠[26] After a conscription law took effect in February 2024, entries from Myanmar into Thailand in March and April were more than three times those of a year before.^⁠[33]

International indices bring these pieces together. Oxford Insights ranks Myanmar 173rd of 195 countries in its 2025 Government AI Readiness Index, below every Southeast Asian country except Timor-Leste.^⁠[34] The IMF's AI Preparedness Index gives Myanmar 0.33, the lowest score in Southeast Asia and about the average for low income countries.^⁠[35]

6Neighbours are building their own

While this happened, Myanmar's neighbours began building AI for their own languages. Singapore announced a S$70 million national programme for multimodal language models in December 2023,^⁠[36] and AI Singapore's SEA-LION models cover eleven languages used in the region, Burmese among them.^⁠[11] In Thailand, SCB 10X released the Typhoon models from late 2023,^⁠[37] and a government backed model, ThaiLLM, opened to the public in 2026.^⁠[38] In Viet Nam, VinAI released PhoGPT, trained from scratch on 102 billion Vietnamese tokens, in 2023.^⁠[39] Indonesia's Sahabat-AI launched in November 2024,^⁠[40] and Malaysia's prime minister launched ILMU in August 2025.^⁠[41]

We found no comparable programme for Burmese. That does not prove there is none, but the Burmese work we could find is small and led by universities and individuals. The Asian Language Treebank was built with the University of Computer Studies, Yangon,^⁠[42] which also produced a Myanmar and English corpus of about 200,000 sentence pairs.^⁠[43] Ye Kyaw Thu released open tools for splitting Burmese text into words,^⁠[44] and developers published community models such as MyanmarGPT and Burmese-GPT on Hugging Face in 2023 and 2024.^⁠[45,46]

The money is moving the other way. Private investment in AI reached 344.7 billion US dollars in 2025, and 285.9 billion of it went to companies in the United States. Singapore, with 1.82 billion, was the only Southeast Asian country among the top 15.^⁠[47] The World Bank reports that high income countries account for 87 percent of notable AI models and 91 percent of AI venture capital, with 17 percent of the world's people.^⁠[48]

7Why the gap compounds

None of this would be urgent if the gap stayed the same size. It does not, because AI improves where it already has users, data and builders, and each of those feeds the others. A language with more text gets better models. Better models attract more users. More users produce more text, and more reasons to build the next model. A language that starts at the bottom of that loop falls further behind each year it stays there.

International institutions describe the same risk. In January 2024 the head of the IMF wrote that many countries "don't have the infrastructure or skilled workforces to harness the benefits of AI, raising the risk that over time the technology could worsen inequality among nations."^⁠[49] A 2025 UNDP report names Myanmar among the countries that "risk missing early dividends" from AI because of gaps in skills, investment, power and connectivity.^⁠[50] Oxford Insights now separates the model makers, the few countries with the resources and talent to build frontier AI, from the model takers.^⁠[34]

Part of the distance can already be measured as a trend. Between 2019 and 2023, Viet Nam's research articles for each one from Myanmar went from about 21 to about 45.^⁠[31] Myanmar's recorded shutdowns went from 37 in 2023, second in the world, to 95 in 2025, first.^⁠[22,51] Between 2023 and 2026, five neighbours put language models of their own into public use. If these trends hold, the gap in 2030 will not be today's gap with a few years added. It will be the gap between countries that shape AI in their own language and a country that waits for others to do it.

The loop can also run the other way, and the evidence shows the gap is not fixed. Tokenizers designed for many languages bring the cost of Burmese down from 11.7 times English to about 1.6.^⁠[20] Google's Gemma 4 31B already scores within two points of Thai in Burmese.^⁠[19] A good open Burmese model makes Burmese products cheaper to build, and those products bring more Burmese speakers online in their own language, which produces the text the next model needs.

8What would change the curve

The work that would change the curve is concrete, and much of it does not need a data centre.

• Burmese text, in Unicode and open. Convert legacy Zawgyi text, clean it, record where it came from, and publish it under open licences, so every model builder starts from the same better data.
• Public Burmese evaluations. Test sets such as Belebele and BURMESE-SAN already exist.^⁠[1,18] Running them on every new model, with native speakers reviewing the answers, turns progress into something measured instead of claimed.
• Models that fit the conditions. Where power and connections fail, a smaller open model that runs on an ordinary computer without a connection is worth more than a larger one in a data centre abroad.
• Burmese speakers as builders. Writing data, judging answers and fixing mistakes needs people who read Burmese, inside Myanmar and abroad, more than it needs hardware.

This is the work Arckara was started to do. Our first model, Arc 1.0, will be open source and is expected in December 2026, and we will publish how it is evaluated, including where it falls short.