The Korean AI startup crossed a major open-source milestone just three months after its first public release.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, announced that its Darwin family of language models has surpassed 1 million cumulative downloads on Hugging Face — reaching over 1.03 million — roughly three months after the model line's initial public release in April 2026. The milestone was driven primarily by community-created derivative and quantized versions of Darwin, rather than downloads of the official model release alone.
VIDRAFT's Darwin language model series has crossed 1 million cumulative downloads on Hugging Face, the Korean AI startup confirmed on July 13, 2026 — a landmark that took just approximately three months to reach from the model's first public debut in April 2026. Total downloads have already exceeded 1.03 million, marking one of the faster community adoption curves seen among Korean open-source foundation models.
The download count is notable not just for its size, but for how it was achieved. According to VIDRAFT, the majority of Darwin's downloads came not from the company's own official repository, but from derivative models that developers around the world independently redistributed or quantized — essentially remixing Darwin for their own use cases. As of the announcement, the Darwin model family spans 269 repositories on Hugging Face, and developers have produced 1,210 distinct quantized variants of the model.
One derivative in particular stands out: a community-created model called Darwin-9B-NEG has individually accumulated more than 560,000 downloads, accounting for a substantial share of the overall milestone.
VIDRAFT also shared performance data for its flagship models. The company says Darwin-398B-JGOS achieved a score of 90.9% on the GPQA Diamond benchmark, a rigorous scientific reasoning evaluation, while Darwin-28B-Opus scored 88.89% on the same benchmark. In South Korea's government-administered K-AI blind evaluation, 11 out of the top 20 ranked models were from the Darwin family. Additionally, Darwin models claimed first place in 14 categories at Polaris, a competitive benchmark focused on predicting molecular properties relevant to drug discovery.
Beyond raw benchmark performance, VIDRAFT highlighted its proprietary inference acceleration technology, branded VKAE, which the company says enables a processing throughput of 601 tokens per second on a single GPU. The system is designed to be deployable via an OpenAI-compatible API, lowering the barrier for developers looking to integrate Darwin models into existing workflows.
The company is also extending Darwin's capabilities beyond text. VIDRAFT recently integrated its own vision-language model (VLM) module into Boston Dynamics' quadruped robot Spot, enabling the robot to respond to voice commands and carry on basic conversations — an early demonstration of the company's ambitions in on-device and physical AI applications.
VIDRAFT CEO Minsik Kim commented on the milestone, saying the 1 million download figure carries special meaning because it reflects choices made voluntarily by developers across the globe. He added that the company will continue to focus on building practically deployable AI models and expanding its presence in the global open-source ecosystem.
Reaching 1 million downloads on Hugging Face is a meaningful threshold in the open-source AI world, and doing so within three months of a first release places VIDRAFT in rare company — particularly among Korean AI companies that develop and publish their own foundation models. The fact that the vast majority of those downloads came from community-derived versions, rather than from VIDRAFT's own releases, underscores a broader dynamic in modern AI development: a model's real-world reach is increasingly measured by how freely the community can adapt and redistribute it.
Darwin's dual support for both Transformer and Mamba architectures has attracted attention from international technical outlets, suggesting the model offers architectural flexibility that developers find compelling. For a startup operating with a comparatively lean infrastructure footprint, these results signal that resource efficiency and open collaboration can be a viable path to global AI influence — even as much larger players dominate the headlines.
The physical AI angle — applying Darwin-derived vision-language capabilities to a real-world robotics platform — also points toward where VIDRAFT sees its next frontier, as the industry increasingly looks to bridge large language models with embodied, sensor-rich systems.
Q: How quickly did VIDRAFT's Darwin model reach 1 million Hugging Face downloads?
A: Darwin crossed the 1 million download threshold in approximately three months, having first been publicly released in April 2026.
Q: What drove Darwin's download numbers on Hugging Face?
A: The bulk of Darwin's downloads came from community-created derivative models — including quantized variants — rather than from VIDRAFT's own official releases. There are currently 1,210 quantized variants of Darwin spread across 269 repositories.
Q: What is VIDRAFT's VKAE technology?
A: VKAE is VIDRAFT's proprietary inference acceleration system, which the company says achieves 601 tokens per second on a single GPU. It is designed for deployment through an OpenAI-compatible API, making it accessible to developers building on standard toolchains.
Source: 중앙일보 (2026-07-13) — original article