Korean AI Startup Vidraft's 'Darwin' Model Surpasses 1 Million Global Downloads
Translated from Korean, summarized and contextualized by DistantNews.
At a glance
- South Korean AI startup Vidraft's language model "Darwin" has surpassed 1 million downloads on the open-source platform Hugging Face.
- This achievement, reached in just three months, highlights the model''s efficiency and effectiveness, even with limited resources.
- Vidraft is now expanding Darwin's capabilities to "on-device AI" applications.
South Korean AI startup Vidraft has achieved a significant milestone, with its language model "Darwin" exceeding 1 million downloads on the global open-source platform Hugging Face. This remarkable feat was accomplished in approximately three months since the initial model's release in April, demonstrating the rapid adoption and interest from the international developer community.
The success is attributed not only to Vidraft's official releases but also to the numerous derivative models created by developers worldwide who have modified or optimized Darwin. The "Darwin" ecosystem now boasts 269 repositories and 1,210 variant models, with one specific derivative, "Darwin-9B-NEG," alone accounting for over 560,000 downloads. The model's technical prowess has garnered attention from international AI communities and Chinese media outlets.
Vidraft's achievement is particularly noteworthy given its status as an early-stage startup in a market increasingly dominated by large corporations like LG, Naver, Upstage, and Kakao, which have also released open-source AI models. Darwin's adoption signifies its ability to compete effectively in the growing market for sovereign AI, where customized models are tailored to national needs.
Darwin's performance has been validated through various benchmarks. Its larger model, "Darwin-398B-JGOS," achieved a 90.9% accuracy rate on the GPQA Diamond scientific reasoning benchmark, while the smaller "Darwin-28B-Opus" scored 88.89%. The model family also ranked highly in government-led K-AI blind evaluations and demonstrated a 99.5% evasion rate for AI trap questions. Vidraft achieved these results using a limited 24-GPU resource environment by employing an "evolutionary methodology" that combines strengths of existing models rather than training from scratch.
Vidraft is now focusing on applying its "VKAE" self-computation acceleration technology to "on-device AI" applications, enabling devices like smartphones and robots to run AI models locally. The company recently integrated its AI module into the "Spot" quadruped robot, enabling voice command and conversational functions. Vidraft CEO Kim Min-sik expressed pride in the community-driven download numbers, stating that Darwin is proving the practical utility of AI models tailored to diverse global environments.
The download numbers, directly from global developers who have used and created them, are very meaningful. Darwin is proving that AI models tailored to each country's environment can be practically utilized in the market.
Originally published by Dong-A Ilbo in Korean. Translated, summarized, and contextualized by our editorial team with added local perspective. Read our editorial standards.