[For more diverse corporate information on the startups mentioned in this article, please visit Unicorn Factory's big data platform 'Data Lab'.]
"Among our clients, many are not large corporations and thus find it difficult to utilize high-spec GPUs (Graphics Processing Units). When combined with top-tier AI (Artificial Intelligence) models, these systems often fail to run smoothly, and securing additional budgets for AI adoption is also challenging. Therefore, we determined the need for lightweight and fast-operating models. Superb AI has focused on developing a lightweight vision AI model capable of delivering high performance without requiring separate additional fine-tuning."
On June 18 last year, at the 'Foundation Few-Shot Object Detection Challenge' held during the main track of CVPR (Computer Vision and Pattern Recognition Conference) 2026—the world's largest computer vision academic conference—domestic startup Superb AI secured first place. Cha Moon-soo, CTO (Chief Technology Officer) and one of the co-founders of Superb AI, established in 2018, explained the background behind developing a model like this as follows: 1Wi Reul (수상).
Superb AI was founded on the 5th, 2018, by five co-founders who joined forces. Their shared experience working together at SK Telecom's T-Brain research institute led to the founding of the company.
In the early stages of their startup, they focused on data labeling automation solutions. Subsequently, in response to client requests, they enhanced their services into an MLOps platform encompassing data collection, management, training, and deployment. They further expanded their portfolio to include proprietary vision foundation models and industry-specific vertical solutions.
CTO Cha stated, "We confirmed that many manufacturing clients lack software engineers entirely, which led us to grow a solution business model that solves problems from start to finish on behalf of the client."
The technology underpinning Superb AI's CVPR victory is its proprietary foundation model named 'Zero.' The name implies a model that operates without prior training. However, this does not mean the model itself was never pre-trained. It more accurately signifies that when applied in real-world scenarios, clients do not need to undergo separate retraining or redeployment processes. In fact, the solution submitted by Superb AI for this competition consisted of a multi-stage pipeline including data and prompt exploration, fine-tuning, and relabeling.
CTO Cha explained, "Vision AI is broadly divided into two approaches: one involves learning predefined classes to achieve high recognition and performance." He added, "If new objects need to be added, the entire process—data creation, retraining, and redeployment—must be reset." He further noted, "The other approach is zero-shot or few-shot methods, where objects are identified by simply inputting class names in natural language or cropping and submitting images, without requiring separate retraining."
Superb AI has concentrated on developing models based on zero-shot and few-shot capabilities. He remarked, "Since we operate in the B2B sector, we determined that lightweight models in limited forms are most suitable for industrial sites." Superb AI developed lightweight models to implement solutions capable of running with minimal memory and computational resources. Additionally, they focused on enabling immediate application across diverse environments without separate retraining.
While Superb AI placed fourth in the same competition last year, this year they climbed three ranks to secure overall first place. CTO Cha stated, "Compared to last year's first-place team, we achieved a significant gap and recorded higher scores, which is meaningful." He added, "To our knowledge, being the first domestic company to win overall first place in an international competition with an industry-specialized vision foundation model is unprecedented."
He explained, "Superb AI's greatest differentiator lies in its long-term accumulation of vision data from industrial sites. The experience of building large-scale datasets and iteratively enhancing models based on them serves as a critical entry barrier."
Capitalizing on this competition victory to validate their proprietary model's performance, the company plans to accelerate its global business expansion. Currently operating local teams in the United States and Japan, they are also considering market entry into Southeast Asia, including Singapore. CTO Cha noted, "Since winning the award, we have seen an increase in overseas companies contacting us directly without any separate sales activities. It seems that our record of winning a certified competition serves as a far more powerful reference than personally explaining our performance."
In the long term, the company intends to expand its existing business axis into the physical AI domain. Rather than developing robot hardware directly, they will focus on serving as a data layer and AI platform that connects various devices such as robots, drones, and humanoids.
He stated, "While some companies are developing robot foundation models themselves, we place greater emphasis on securing platforms and technologies that connect diverse hardware atop these models." He continued, "Taking autonomous driving as a reference, it can be categorized into perception, reasoning, and action. Currently, our strongest competitive area is perception, and we are also developing technologies to bridge reasoning and action."
[MoneyToday startup media platform Unicorn Factory]
in FUTURE
