LG AI Research signs Kazakhstan to build a Kazakh-language EXAONE

Building the Kazakh data is the expensive part of this agreement. Kazakh did not register at all in the Hugging Face dataset statistics the World Bank compiled for its World Development Report 2026 ^[8]. On those numbers there is no public corpus to fine-tune from, so the collection work sits inside the deal, and LG AI Research agreed to run the joint development with Kazakh companies and universities ^[4]. The institute did not disclose a contract value ^[19].

Put together the three blocs a sovereign AI buyer is trying not to depend on and you reach 92.8% of the registered datasets, counting other European languages at 26.3% alongside English and Chinese ^[7][9][1]. Arabic is 2.4% and Korean 2.2% ^[10][11]. Everything else on the list shares 2.6% ^[2]. Korean's share is roughly a twenty-eighth of English's ^[3], and Arabic's sits 0.2 points above Korean's ^[4], so the Middle East sovereign wealth fund LG hopes to sign next month ^[5] has marginally more of its own language in that data than the institute offering it independence from US and Chinese models ^[1]. These are counts of registered datasets, not tokens.

The paying customer named for this product line is the London Stock Exchange Group, signed last September, with the Middle East rollout to follow a proof-of-concept phase ^[6]. Two memoranda and one exchange group is the commercial base under the export push. It is also the first time the institute has pursued joint development or commercialization with non-English-speaking emerging economies ^[17].

EXAONE 4.5, unveiled in April, outperformed OpenAI's GPT-5 mini and Anthropic's Claude Sonnet 4.5 on science, technology, engineering and mathematics benchmarks ^[12]. One of those two comparators names itself a mini. Since then the institute has shipped EXAONE BI, the manufacturing-focused Tabula and Discovery for materials research ^[13], taken on a next-generation model built on Nvidia's Nemotron platform under the group's alliance with Nvidia ^[14], and joined a government project upgrading K-EXAONE into an independent national foundation model ^[15]. The Kazakh work competes with all of that for the same engineers.

Two readings fit the evidence. In the first, the memoranda convert, Kazakhstan puts the model into industry, energy, healthcare and logistics ^[4], and LG has a line of business in countries where big tech models have not properly learned the language ^[18]. In the second, the memoranda stay diplomatic, and NAVER's Thai and Arabic models, along with SK Telecom's and Upstage's partnerships ^[16], bid the same pitch down before much of it is billed. A third path cuts against the vendor: once a Kazakh corpus exists it is portable, and the next lab to fine-tune on it does not pay for the collection. I would weight the first higher in the Gulf, where a proof-of-concept and a named exchange client already exist ^[6], and lower in Kazakhstan, where the deliverables named so far are joint development and talent training ^[4].