The archive · AI & Models · Technical decision · 2024–2026
Trillion Labs builds Korea's own LLM from scratch; three big models in a year
Ex-Naver HyperCLOVA researcher Shin Jae-min founded Trillion Labs in 2024 to build a Korean LLM from scratch; a $4.2M pre-seed and 3 models in a year followed.
Trillion Labs (트릴리온랩스)
What the business is
A Seoul AI startup pre-training Korean-native large language models from scratch (data collection to pre-training to verification), with multilingual and reasoning-focused models plus reproducibility research.
Starting capital：About $4.2M (₩5.7B) pre-seed announced Sep 2024, led by Strong Ventures with Kakao Ventures, Base Investment, The Ventures, Goodwater Capital and BAM Ventures; later up to $1M in AWS generative-AI accelerator credits.
How it started
Trillion Labs was founded in Seoul in 2024 by Shin Jae-min, shortly after the sovereign-AI debate made Korean-language models a strategic question. It announced a $4.2M pre-seed in Sep 2024 from Strong Ventures, Kakao Ventures and others, planning to finish a Korean-native foundation model by end of that year.
What happened
In March 2025 it released Trillion-7B, a multilingual model trained for about ₩150M; in July, Tri-21B claimed to beat LLaMA and Qwen on math, coding and reasoning; in August it released Tri-70B with full training checkpoints — a first for a Korean startup — betting on reproducibility. In October it became the only Korean LLM startup picked for AWS's generative-AI accelerator (up to $1M in credits) and published rBridge, a method that predicts a 32B model's performance from a ≤1B one, cutting evaluation cost by about 100x.
How it ended up
Still building: selected for South Korea's MSIT AI foundation-model project in the Lunit medical-science consortium, it is leading development of a Co-Scientist foundation model for genomics and clinical data, targeting a ₩140T medical-AI market.
Background
Shin Jae-min helped build Naver's HyperCLOVA X before deciding Korea needed something more radical. In 2024 he founded Trillion Labs in Seoul to pre-train Korean-native large language models from scratch — data collection, pre-training and verification fully under the company's control — rather than fine-tuning Western open models. He argued that a business built on 'someone else's goodwill' is not sustainable and that only owning the whole pipeline enables advanced agents.
The bet attracted capital early: a $4.2M pre-seed in Sep 2024 led by Strong Ventures with Kakao Ventures and others. By March 2025 Trillion Labs shipped Trillion-7B, a multilingual model trained for about ₩150M; in July, Tri-21B claimed to beat LLaMA and Qwen on math, coding and reasoning; in August it released Tri-70B with all training checkpoints public — a first for a Korean startup — making reproducibility a core promise.
In October 2025 it became the only Korean LLM startup selected for AWS's generative-AI accelerator, securing up to $1M in cloud credits, and published rBridge, a technique predicting a 32B-parameter model's performance from a sub-1B one, claiming ~100x cheaper evaluation. It was then chosen for South Korea's MSIT AI foundation-model project in a medical-science consortium led by Lunit, where it leads a Co-Scientist foundation model for genomics and clinical data — still early, but with a credible from-scratch track record.
What has to be true
- From-scratch training became the moat: while Korean startups fine-tuned LLaMA or Qwen, Trillion Labs owned data, pretraining and verification — the only sustainable basis for advanced agents.
- Reproducibility as marketing: releasing full checkpoints for a 70B model made openness itself a differentiator and won technical credibility in a market used to 'fake-local' models.
- Cost engineering won supporters: a ~₩150M training bill and an rBridge method that cuts evaluation cost ~100x showed sovereign AI need not require US-scale budgets.
- Sovereign-AI timing: the 2024-2025 global debate over national AI independence gave a young Korean lab a strategic reason to exist — and partners to prove it.
What can be applied
From-scratch training with open checkpoints gave Trillion Labs credibility — reproducibility and independence — while rivals fine-tuned models.
Aftermath
As of 31 Aug 2026, Trillion Labs is still building rather than scaling: its latest public milestones are the Oct 2025 AWS accelerator selection and rBridge, and selection for the MSIT AI foundation-model project, where it leads the Co-Scientist medical-science model with Lunit. No revenue or Series A has been publicly announced; the company's next tests are turning open-checkpoint credibility and national-project participation into commercial adoption.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card