Nums AI Adds KRW 2B in Second Pre-Seed Funding Close

July's first investors are being joined by two more firms at Nums AI, as the Seoul startup builds a business around prediction from structured data. Mirae Asset Venture Investment and Korea Investment Partners have supplied KRW 1B each in a second pre-seed closing, according to November 4 reporting from beSUCCESS. The new KRW 2B brings cumulative venture investment to KRW 6B, extending the financing behind models intended to answer questions from tables and graphs.

Co-founder and CEO Jaemin Yoo, a Seoul National University computer-science professor, and fellow co-founder Dooho Lee are working in a category with a demanding customer brief: make predictions useful across different datasets while reducing the need to build a new model for each assignment. Nums AI plans to put the funding into model development and hiring. For an enterprise buyer, the interesting consequence is how much engineering work could move from training another model to deciding which prediction belongs in an actual business process.

A second close extends Nums AI's pre-seed financing

The financing has a clear sequence. Nums AI's July 20 first-round announcement reported KRW 4B led by Stonebridge Ventures, with SBVA, KT Investment and BASS Ventures participating. The second close adds KRW 2B to that base. Describing the latest event as KRW 6B of newly raised capital would erase the earlier financing and exaggerate the incremental money arriving now.

Korea Startup Post's account also reports KRW 800M in TIPS R&D support, separate from the investment. That support adds a research resource without changing the KRW 6B venture total. The reports identify a February 2026 founding by researchers from Seoul National University and KAIST; they do not disclose a valuation or the precise date on which the second close legally settled.

This combination gives the company two kinds of work to fund. Improving a foundation model requires research and experimentation, while delivering it to customers requires inference software, deployment choices and people who can translate a customer's problem into a usable prediction task. The funding announcement names model advancement and recruitment as priorities. Any revenue contribution from that work remains undisclosed, so the financing should be read as a commitment to development rather than a measure of commercial scale.

Causilo makes examples part of the prediction interface

Nums AI's Causilo introduction describes a pretrained model that receives rows with known outcomes and new rows whose outcomes need to be predicted. The first set provides context for the second. Causilo uses that context for classification or regression while keeping its pretrained weights fixed, avoiding a separate weight-training exercise for each new task.

The practical change is in how a prediction problem reaches the model. A customer supplies examples that establish the relationship being investigated, such as records paired with an observed outcome. The model then predicts the missing outcomes in the query rows. This makes the table itself part of the interface, which is a more concrete proposition than asking a business to buy an abstract layer of artificial intelligence.

Nums AI's technical report places Causilo within research on tabular foundation models and describes refinements before compressing rows into representations used for prediction. That work reflects an engineering trade-off between retaining information and managing computation. It gives a technically minded buyer something to inspect beyond a funding headline, while leaving the buyer responsible for comparing performance on the actual task and data it intends to use.

Research access and production use have different terms

The route from research evaluation to a commercial deployment includes a licensing decision. Nums AI provides public code and downloadable model weights, but the weights carry a separate Causilo license. The official terms distinguish non-commercial research from commercial or production use requiring separate licensing. Calling the whole package unrestricted open source would give a prospective buyer the wrong purchasing assumption.

The Causilo product page also points toward different ways of building with the technology. For a customer, the implementation question extends beyond obtaining a checkpoint: where will the model run, how will data reach it, and who will maintain the prediction workflow? These are procurement and operating questions that follow from the product's stated deployment options, rather than proof that any particular customer has already resolved them.

The company's documentation also asks users to evaluate on held-out outcomes and compare against a suitable baseline. That advice belongs close to the product promise. A pretrained model can change the amount of task-specific model-building work while still needing representative examples, appropriate evaluation and a decision about the cost of being wrong. A forecast becomes commercially useful inside those constraints, with somebody accountable for acting on it.

Hiring connects foundation-model research to delivery

Nums AI's current careers page makes the hiring priority more tangible. Research Scientist work includes architectures, synthetic tasks and in-context learning; ML Engineer work includes inference infrastructure and deployment. The roles are based in Seoul. Together they describe a team expected to connect experimental progress with software that another organization can use repeatedly.

Nums AI also develops Ephris for graph learning, according to its official company overview. That broadens the structured-data research agenda beyond tables, but it should not be mistaken for disclosed revenue diversification or an established customer footprint. The public record reviewed for this announcement provides no verified revenue, paid-customer total or valuation, and illustrative use cases are not a substitute for those figures.

The additional investors are therefore joining a research-led business while much of its commercial record is still ahead of it. For Jaemin Yoo, Dooho Lee and the people Nums AI recruits, the work now runs across model quality, inference systems and the customer conversation about where a prediction fits. Each new dataset brings another opportunity to connect the examples a company already holds with a decision its people still have to make.