The language model “Darwin” product family, developed by Korean artificial intelligence (AI) startup VIDRAFT, has surpassed 1 million cumulative downloads on the global open-source platform Hugging Face. This milestone, achieved in about three months since the first model was released in April this year, corresponds to an actual measured cumulative download count of 1,031,761.

This performance was driven mainly by derivative models, not only the official models distributed directly by the company, but also models that developers around the world voluntarily adopted, modified, or compressed and then redistributed. There are 269 repositories related to Darwin and 1,210 variant models, and a specific derivative model (Darwin-9B-NEG) alone has recorded more than 560,000 downloads. In particular, overseas AI communities and Chinese media have highlighted Darwin’s technical characteristics.

Among Korean companies that have released their own AI models as open source and exceeded 1 million downloads, major conglomerates and mid-sized companies such as LG, Naver, Upstage, and Kakao are representative. Despite being a relatively small early-stage startup, VIDRAFT has secured a meaningful adoption rate in a market where competition for country-tailored AI (sovereign AI) is intensifying, through the voluntary choice of developers worldwide.

Darwin’s performance is demonstrated across multiple indicators. In the GPQA Diamond benchmark, which evaluates scientific reasoning capability, the large-scale model “Darwin-398B-JGOS” achieved an accuracy of 90.9%, while the smaller “Darwin-28B-Opus” recorded 88.89%. In the government-led K-AI blind evaluation, 11 of the top 20 models were found to be from the Darwin series. Darwin has also secured first place in 14 categories of the new drug property prediction competition (Polaris), and a 99.5% avoidance rate for its own AI trap questions.

These outcomes were achieved using a limited resource of 24 graphics processing units (GPUs). Rather than training models from scratch, VIDRAFT adopted an “evolutionary methodology” that enhances efficiency by combining the strengths of existing models and evolving them into the next generation.

In addition, by applying its proprietary computation acceleration technology “VKAE,” the company has enabled processing of 601 words (tokens) per second on a single-computer environment, representing a roughly 1.5 to 2-fold speed improvement over previous levels. VIDRAFT is expanding this technology into the “on-device AI” field, in which AI runs directly on devices such as smartphones and robots. Recently, it equipped the quadruped robot “Spot” with its own AI module to implement voice command and conversational functions.

VIDRAFT CEO Kim Min-sik stated, “It is highly meaningful that this download figure was generated by developers around the world who used the models themselves,” adding, “Darwin is demonstrating that AI models tailored to each country’s environment can be practically and usefully applied in the market.”

Popular News