Korea Investment Corp Runs Staff AI on Alibaba’s Qwen3 Model

Yonhap reported on 19 October 2026 (Seoul time) that Korea Investment Corporation (KIC), the sovereign wealth fund that manages 339 trillion won, runs its in-house AI on Chinese-developed models. Its language model is Alibaba’s Qwen3, and its search embedding model is BGE-M3 from the Beijing Academy of Artificial Intelligence (BAAI).

The figures come from materials KIC submitted to Kang Min-guk, a People Power Party lawmaker on the National Assembly’s Finance and Economy Planning Committee, Yonhap said. According to the report, BAAI was placed on the US Commerce Department’s export-control list in March of last year.

Business Pill · OPEN WEIGHTS

A short explainer on open weights: renting a model through an API or running the weights yourself on your own hardware. It teaches the general idea only and says nothing about any company in this story.

What KIC Put Inside Its In-House AI

According to Yonhap, the in-house AI carries four models. Besides Qwen3 and BGE-M3, they are Meta’s Llama3, which blocks harmful content, and Google’s Gemma3, which handles images. Qwen3 was the only one KIC classified as a language model, the model that understands documents and generates answers.

Yonhap said this means document summaries and translations, as well as work handling internal rules and weekly reports on the groupware and lawmakers’ question materials in the National Assembly’s work-management system, all pass through Chinese-developed models.

KIC has offered the service to all 350 employees since 7 May 2026, Yonhap reported. It added BGE-M3 to the in-house AI on 15 December of last year.

The Export-Control Question

Yonhap reported that BAAI was added to the US Commerce Department’s export-control list on 25 March of last year, on the grounds that it supported China’s military modernisation.

KIC explained, according to Yonhap, that it chose its models on accuracy, comprehension, reasoning, fit for its work and stability, and that a limited hardware set-up of four GPUs was a key factor. On its evaluation, KIC said Qwen3 came out ahead on cost-performance, spare VRAM and suitability for Asian languages.

Kang’s point, as Yonhap reported it, is that the evaluation criteria KIC submitted did not include the developer’s nationality or whether it is subject to US export controls.

The Use Numbers

Yonhap reported that use of the in-house AI fell 78.7% in four months, from 3,755 uses in May, the launch month, to 798 in September. Users fell 56.6% over the same period, from 339 to 147.

KIC employees used ChatGPT 36,881 times last month, about 46 times the in-house AI, according to the report. KIC said the in-house AI is used in a limited way, to assist with nine kinds of tasks.

On the same day it launched the service, 7 May 2026, KIC signed a contract worth 950.4 million won to expand related infrastructure, Yonhap said; the project is due to finish next month and take the GPU count from four to twelve. KIC said the expansion supports an internal AI platform upgrade ordered in August, and that it placed the order with rising GPU and memory prices and delivery times in mind.

What the Lawmaker Is Asking For

Kang said, according to Yonhap: “성능과 가성비는 따지면서 개발사 국적과 제재 여부는 기준에도 없었고, 사내 AI 이용이 5분의 1로 줄어드는데 장비를 3배로 늘렸다”. In our translation: performance and cost-effectiveness were weighed, but the developer’s nationality and sanctions status were not even among the criteria, and the equipment was tripled while in-house AI use fell to a fifth.

He called on KIC to check that the added equipment is used in line with actual demand, and on the Ministry of Finance and Economy, as the supervising ministry, to review how the public bodies under it adopt AI models and to set nationality and security criteria, according to Yonhap.

The Structural Reading

As we read it, the materials describe a familiar trade-off at a public institution: with four GPUs, KIC picked the model that scored best on cost-performance and Asian-language fit, and the selection sheet, as Kang describes it, did not ask who built the model.

As we read it, the question Kang raises is less about any one model than about procurement rules: which criteria a state investor applies when the open-weight models it adopts include one, BAAI’s BGE-M3 embedding model, from a developer on a US export-control list. We covered China’s proposed export rules for AI technologies last week, and the spread of Chinese open-weight models is a recurring theme in our coverage.

The Structural Read

As we read it, the hardware detail matters: KIC explained, according to Yonhap, that a limited set-up of four GPUs was a key factor, and that Qwen3 came out ahead on cost-performance, spare VRAM and Asian-language fit.

As we read it, the export-control point attaches to one model, not all of them: Yonhap reported that BAAI, the developer of the BGE-M3 search model, was put on the US Commerce Department’s export-control list on 25 March of last year.

As we read it, the use numbers frame Kang’s budget argument: Yonhap reported in-house AI use down from 3,755 times in May to 798 in September while the GPU count is set to go from four to twelve, and staff used ChatGPT 36,881 times last month.

Three Implications

PUBLIC-SECTOR AI BUYERS Kang called on the Ministry of Finance and Economy to review how the public bodies under it adopt AI models and to set nationality and security criteria, according to Yonhap.

OPEN-WEIGHT MODEL DEVELOPERS KIC said Qwen3 came out ahead on cost-performance, spare VRAM and suitability for Asian languages, Yonhap reported.

IN-HOUSE AI AGAINST PUBLIC CHATBOTS Yonhap reported that KIC staff used ChatGPT 36,881 times last month, about 46 times the in-house AI.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework puts it this way: “Every technology stack has layers. Some are scarce, some are abundant.”

As we read it, KIC treated the model as the abundant layer, picking the open-weight option that fit its hardware and budget, while the questions Kang raises, about who built a model and whether its developer is under export controls, sit in a layer the selection sheet did not price. The framework was published in March 2026 and does not discuss KIC.

What Is Not Established

We read Yonhap’s report of 19 October 2026 in full. We did not read the materials KIC submitted to Kang Min-guk, or any statement KIC published itself; KIC’s explanations above are as Yonhap reported them. The report does not say that any data left KIC’s systems or that any security incident occurred, and none of the texts we read says whether KIC plans to change its models.

The Bottom Line

Korea Investment Corporation, which manages 339 trillion won, runs its in-house AI with Alibaba’s Qwen3 as the language model and BAAI’s BGE-M3 for search, Yonhap reported on 19 October 2026, citing materials submitted to lawmaker Kang Min-guk. Use of the in-house AI fell from 3,755 times in May to 798 in September, while staff used ChatGPT 36,881 times last month, according to the report.