ZERO is a ready-to-use vision foundation model built for industrial use.
It doesn’t just see what it was trained on—ZERO sees, searches, and understands exactly what you need.
Higher detection accuracy with text prompts — fewer missed detections
Better understanding of Korean and descriptive prompts — like “orange jacket”
More distinct confidence scores — easier threshold tuning
Vision AI expertise built on real-world industrial data — proven with first place in the few-shot object detection challenge at CVPR 2026, the world’s largest computer vision conference.
53.9mAP
2nd-place score: 51.6 · Organizer baseline: 33.3
5
First place in 5 of 7 evaluation categories
1st
From 4th in 2025 to 1st in one year
1Define Problems
2Collect Data
3Label Data
4Train a Model
5Deploy
1Define Problems
2Deploy
ZERO is a ready-to-use vision foundation model built for industrial use.
It doesn’t just see what it was trained on—ZERO sees, searches, and understands exactly what you need.
| ZERO | Conventional AI | Conventional SI | |
|---|---|---|---|
| Instant adaptation to new requirements | Only recognizes what was pre-trained | 2-6 months for update | |
| Flexible business expansion | Additional development and costs required | ||
| Sustainable and efficient management | Increased overhead from siloed solutions | Maintenance costs incurred |
Text Prompt
Accurately find the objects you describe in natural language.
Image Prompt
Detect objects intuitively — upload an image or specify with boxes and points.
From accuracy-critical inspection to real-time processing on edge devices. Whichever model size you choose, text and visual prompting work the same way.
| Model | Deploy parameters | Active parameters at inference | Recommended for |
|---|---|---|---|
| ZERO-Tiny | 440M | 65M | Edge · real-time processing |
| ZERO-Base | 500M | 125M | General purpose |
| ZERO-Large | 610M | 235M | Peak accuracy |
ZERO-Tiny runs real-time inference at 0.329 TFLOPs on 640×640 input, keeping most of the larger models’ detection performance while cutting operating costs in production.
SPBEval Text-G / Visual-G AP · at 640×640
| Model | Active parameters | Compute | Text AP | Visual AP |
|---|---|---|---|---|
| ZERO-Tiny | 65.0M | 0.329 TFLOPs | 29.7 | 23.7 |
| YOLOE-26x | 69.9M | 0.197 TFLOPs | 14.2 | 15.4 |
More than double the detection accuracy of similarly sized real-time detection models. Prompt caching keeps inference running in real time.
Measured on SPBEval, our industrial-domain benchmark — see the latest version’s benchmark performance and the gains delivered with each release.
ZERO 2.2 vs Open models
SPBEval Text AP · ZERO 2.2 (Large)
SPBEval — Superb AI’s in-house industrial-domain benchmark · Visual AP charts follow the same format (ZERO 2.2 Visual-G 26.2 / Visual-I 56.3)
Improving with every version
SPBEval Text AP · trend by version
What are ZERO’s key performance advantages?
• Accuracy — top-tier detection accuracy among open models on SPBEval, Superb AI’s in-house industrial benchmark (ZERO 2.2 Text AP 31.3 vs. 20.6 for the next-best model)
• Compute efficiency — the lightweight model delivers real-time inference with 65M active parameters at 0.329 TFLOPs (640×640)
• Scalability — three sizes (Tiny / Base / Large), with the same prompting workflow from real-time edge inference to high-precision inspection
Can I deploy ZERO through AWS?
Yes. ZERO 2.2 is available on AWS Marketplace today. Deploy it as an Amazon SageMaker endpoint and call it via API — if you already run on AWS, you can adopt it with your existing AWS account; no separate contract required.
What’s the biggest difference between ZERO and conventional AI?
ZERO is Superb AI’s innovative vision foundation model (VFM) that detects and analyzes objects or situations with simple prompts—text, boxes, sketches, or more—without any model training. Unlike traditional vision AI, which requires months of data collection, labeling, training, and deployment, ZERO skips the complexity, making AI adoption faster, more affordable, and accessible to anyone.
Do I need AI expertise or manual labeling to use ZERO?
Not at all. VFM ZERO is designed to be accessible even to non-experts. With intuitive prompt-based interaction, you can use the model right away—no coding experience or large-scale data labeling required.
What types of objects or situations can ZERO detect? Can it recognize new targets it hasn’t been trained on?
ZERO isn’t limited to pre-defined categories. It can detect virtually any object, action, or situation based on your prompt—thanks to its powerful zero-shot capabilities. Even if the model has never been explicitly trained on the target object or scenario before, it can still identify it through your natural language description.
How long does it take to implement ZERO in real-world applications, and is it compatible with our existing video systems?
One of ZERO’s biggest strengths is instant deployment. Since no training is required, you can get started right after prompt setup and system integration. ZERO also works seamlessly with most existing video equipment that supports standard protocols—including CCTV, IP cameras, and drones—especially when deployed through Superb AI’s Video Analytics solution.
