TL;DR
On 30 October 2026, five Korean outlets reported on VIDRAFT's recent work: Chosun Biz on the ViSQL text-to-SQL engine, AI Times on the demo that Hugging Face's open-source team built for our decision AI, and Economist (Korea) on QuantID's analysis of the official Hugging Face benchmark leaderboards. Financial News covered the Chitos security engine, and JoongAng Ilbo covered pretraining a language model on a Korean inference NPU.
| Outlet | Topic | Link |
|---|---|---|
| Chosun Biz | ViSQL: ask in words, get verified SQL | Read |
| AI Times | Hugging Face open-source team builds and donates a demo for Darwin-35B-ZTC-Vision | Read |
| Economist | QuantID: 53 official Hugging Face benchmarks, Korea second in #1 spots | Read |
| Financial News | Chitos: modules covering all 17 FSI AI agent security evaluation criteria | Read |
| JoongAng Ilbo | Pretraining a language model from scratch on a Korean inference NPU | Read |
ViSQL (Chosun Biz)
ViSQL turns a question in plain language into SQL, runs it on a read-only database, checks the result, and writes tables, charts and a report. Darwin-180B-NL2SQL answered 419 of 500 questions on BIRD Mini-Dev (SQLite) and the result was submitted to the BIRD team for listing. The public demo includes a voice-phishing fund-tracing case that drafts a statutory notice from query results.
- Demo: final-bench-nl2sql-demo.hf.space
- Our post: 83.80 on BIRD Mini-Dev
Hugging Face demo (AI Times)
Hugging Face's open-source team built a hands-on demo for Darwin-35B-ZTC-Vision, our vision decision model, and donated it. You give it an image and options, and it returns a probability for each option instead of a long text answer.
Benchmark analysis (Economist)
QuantID counted 53 official Hugging Face benchmarks with 48 leaderboards on 29 October. By the publisher's country, China holds 24 #1 spots, Korea 15 and the USA 2. All 15 Korean #1 spots are FINAL-Bench (VIDRAFT) entries, scored as publisher-reported on Hugging Face. QuantID is a technology partner of VIDRAFT.
Chitos security engine (Financial News)
Chitos, our AI agent security engine, now has modules that address all 17 evaluation criteria in the Financial Security Institute's AI agent security framework. It also runs fully on premises, for networks with no internet access.
NPU pretraining (JoongAng Ilbo)
JoongAng Ilbo reported that VIDRAFT pretrained VKORE-325M-NPU-Base (325.6M parameters) from random initialization on Rebellions ATOM-Max, a Korean NPU that was released for inference. To compare fairly, we trained the same model on an NVIDIA B200 GPU with the same initial weights, data, batch size and learning-rate schedule: 76,292 steps on FineWeb-Edu, about four days, with no divergence. The weights and evaluation scripts of both models are public.
Important: this works only when VIDRAFT's own kernel is installed on the NPU. The training method is patent pending.
Top comments (0)