A lean Korean AI startup just outpaced much larger rivals on the world's biggest open-source model platform.
TL;DR: VIDRAFT, a South Korean AI startup, announced that its Darwin family of language models surpassed one million cumulative downloads on Hugging Face approximately three months after the first model was released in April 2026. The total recorded download count stands at 1,031,761, driven largely by derivative models that developers worldwide built on top of the Darwin base. VIDRAFT achieved this milestone with significantly constrained hardware resources, using an evolutionary model-development methodology rather than training from scratch.
VIDRAFT's Darwin AI model family has crossed one million global downloads on Hugging Face, making the South Korean startup one of a small group of Korean companies — alongside much larger names like LG, Naver, Upstage, and Kakao — to reach that threshold with an open-source language model. The milestone was reported by 동아일보 on July 13, 2026, roughly three months after Darwin's public debut.
The verified cumulative download count for Darwin reached 1,031,761 on Hugging Face, the world's leading open-source AI model platform. What makes that number particularly notable is where those downloads came from: the majority were not of VIDRAFT's official releases, but of derivative models that independent developers around the world forked, fine-tuned, quantized, and redistributed on their own.
The Darwin ecosystem on Hugging Face now spans 269 repositories and more than 1,210 distinct model variants. A single community-built derivative, Darwin-9B-NEG, alone accounts for over 560,000 downloads — more than half the total — a strong signal that the global developer community has found the underlying architecture genuinely useful rather than just novelty-driven.
Darwin's benchmark performance adds weight to that developer enthusiasm. On the GPQA Diamond evaluation, which tests scientific reasoning, the large Darwin-398B-JGOS model scored 90.9 percent, while the lighter Darwin-28B-Opus achieved 88.89 percent. In South Korea's government-run K-AI blind evaluation, 11 of the top 20 ranked models were from the Darwin family. VIDRAFT also reported a first-place finish across all 14 categories of the Polaris drug-property prediction competition, and a 99.5 percent avoidance rate on the company's own AI "trap" question set designed to probe for hallucinations or manipulation vulnerabilities.
All of this was accomplished with a constrained hardware footprint. Rather than training a new model entirely from scratch — a process that typically demands enormous compute — VIDRAFT developed what it describes as an evolutionary methodology: identifying the strengths of existing models and systematically combining and advancing them across generations. The company also developed a proprietary inference acceleration technology called VKAE, which enables processing of 601 tokens per second on a single machine, a speed the company says is roughly 1.5 to 2 times faster than conventional approaches.
Beyond server deployments, VIDRAFT is extending Darwin and VKAE toward on-device AI — running models directly on edge hardware such as smartphones and robots. The company recently demonstrated this by integrating its AI module into a Boston Dynamics Spot quadruped robot, enabling voice command and conversational interaction on the device itself.
CEO Kim Min-sik was quoted by 동아일보 as saying the download figures carry special meaning precisely because they reflect developers choosing Darwin of their own accord, and that Darwin is demonstrating in practice that AI models tailored to different national environments can deliver real-world value globally.
VIDRAFT's Darwin milestone is significant for several reasons that go beyond raw download counts. The sovereign AI movement — the push by individual countries and regions to build AI infrastructure that reflects local language, law, and cultural context rather than relying solely on American or Chinese models — is accelerating. VIDRAFT is competing in that space as an early-stage startup, yet it is drawing organic global adoption that previously belonged almost exclusively to well-capitalized Korean conglomerates.
The organic derivative ecosystem is arguably the most telling metric. When thousands of developers worldwide independently take a model, modify it, and share it again, it suggests the base model offers something architecturally sound and practically flexible. For a startup operating with limited resources, that kind of community-driven distribution is both a validation signal and an effective, cost-free scaling mechanism.
The on-device push also positions VIDRAFT at the intersection of two fast-growing markets: edge AI and embodied intelligence (robots and physical AI agents). Successfully running competitive language models on resource-constrained hardware — demonstrated concretely with the Spot robot integration — suggests the company's efficiency-first philosophy extends well beyond benchmarks.
Q: What is VIDRAFT's Darwin AI model and where can it be downloaded?
A: Darwin is a family of open-source language models developed by South Korean AI startup VIDRAFT, available on Hugging Face. As of July 2026, it has accumulated over one million downloads across official releases and community derivative models.
Q: How did a small startup achieve competitive AI benchmark results with limited resources?
A: VIDRAFT uses an evolutionary model development methodology — combining and advancing strengths from existing models rather than training entirely from scratch — alongside a proprietary inference acceleration technology called VKAE, enabling high performance within a constrained hardware budget.
Q: What is VIDRAFT doing with on-device AI?
A: VIDRAFT is applying its Darwin models and VKAE acceleration technology to edge hardware, including smartphones and robots. The company recently demonstrated the capability by deploying its AI module on a Boston Dynamics Spot quadruped robot, enabling voice commands and real-time conversation directly on the device.
Source: 동아일보 (2026-07-13) — original article