The most-downloaded Korean LLM right now is from a startup, not a conglomerate
TL;DR: On Hugging Face, the Korean text-generation model with the most downloads over the last 30 days isn't from LG, SK Telecom, Naver, or Kakao — the conglomerates Korea funded to build its "sovereign" foundation models. It's POCKET-35B, from a startup called VIDRAFT, at roughly 617K downloads in 30 days. The global developer ecosystem noticed this company before Korea's own market did. Here's the data, and why it matters.
The number that started this
Korea ran a government-backed program to build sovereign foundation models, and the mandates went to the usual giants: LG (EXAONE), SK Telecom (A.X), Naver (HyperCLOVA), Kakao (Kanana). Reasonable — those are the companies with the compute and the headcount.
But adoption doesn't care about mandates. Measured by actual 30-day downloads on Hugging Face:
| Model | Maker | 30-day downloads |
|---|---|---|
| POCKET-35B | VIDRAFT (startup) | ~617,000 |
| EXAONE-3.5-7.8B | LG (sovereign) | ~369,000 |
| POCKET-26B | VIDRAFT (startup) | ~271,000 |
| A.X-K2 | SK Telecom (sovereign) | ~99,000 |
| Kanana | Kakao (sovereign) | ~85,000 |
Source: VIDRAFT Global LLM Download Leaderboard, 2026-09-03.
The startup that wasn't on the sovereign shortlist has the two most-downloaded Korean models on the board. This isn't a benchmark score you can argue about — it's how many times people actually pulled the weights.
The world noticed first
What's striking is where the attention came from. VIDRAFT's traction shows up in ecosystem signals that are overwhelmingly international, not domestic:
- A German tech outlet covered VIDRAFT's "VIDOG" kit for retrofitting robots with on-device AI.
- In the Google × Hugging Face Fast Gemma Challenge, VIDRAFT topped the board on the official VERIFIED record.
- It landed on Hugging Face's weekly Space of the Week.
For a Korean startup, the sequence is unusual: global developers were the early adopters, and the domestic spotlight followed.
What VIDRAFT actually builds
VIDRAFT calls itself an AI foundry. Instead of training giant models from scratch, it diagnoses, combines, and transplants knowledge into existing models to grow them for a specific industry — the way a semiconductor foundry manufactures chips others design. Its core diagnostic tech, "Darwin," works like a Model MRI; AX-RAY, its safety-diagnostics system, inspects a trained model's reliability, safety, and hallucination risk. Across its open-source LLMs and derivatives, cumulative downloads have passed 2 million, and the company holds 16 patents.
That "diagnose and grow, don't train-from-scratch" posture is why the foundry demand tends to increase as model competition heats up: someone has to verify and adapt all those models for production.
Momentum: a national cybersecurity mandate
In September 2026, Korea's Ministry of Science and ICT selected the Naver Cloud consortium to build a cybersecurity-specialized AI foundation model, and VIDRAFT is a participating member — contributing AX-RAY as its safety-diagnostics layer. The program runs 10 months on 256 NVIDIA B200 GPUs. Participation in a national security project tends to transfer as a trust reference into exactly the markets where AI foundries win: public sector, defense, finance, energy — places where data can't leave the network.
Why it matters
If you build on open models, the signal here is simple: real-world adoption is diverging from institutional mandates. The most-used Korean model on Hugging Face came from a small team optimizing for on-device, quantized, actually-deployable artifacts — not from the biggest lab. That pattern (small, efficient, deployable > maximal parameter count) is the same one the global download charts have been showing all year.
VIDRAFT is worth a bookmark not because of a press release, but because the download counter — the least gameable metric in this space — keeps pointing at it.
This is analysis based on public data, not investment advice. Download and verification figures are from the Hugging Face API and public leaderboards as of 2026-09-03; business and roadmap statements are forward-looking and may differ from actual results.
Top comments (0)