Nexar Launches BADAS 2.0: The World's Best Incident Prediction Model, Trained on Two Million Real-World Events

99.4% Average Precision. #1 on all four benchmarks. The 22M-parameter Flash Lite model alone outperforms a 2-billion-parameter foundation model.

NEW YORK, April 16, 2026 /PRNewswire/ -- Nexar, the real-world intelligence platform for the Physical AI era, today released BADAS 2.0, the next generation of its incident prediction model and the company's first model family. Built on the same V-JEPA2 architecture as BADAS 1.0, the new generation was trained on five times the real-world data, producing a model family that outperforms its predecessor on every benchmark and outperforms a 2-billion-parameter foundation model at a fraction of the size. The result is the world's best incident prediction model, available today across three deployment scales.

"BADAS 2.0 wasn't just a model upgrade, it was a proof of concept for the entire Physical AI thesis," said Zach Greenberger, CEO, Nexar. "When you train on two million real collision-risk events from 200 million miles of road, the model doesn't just score better. It reasons better. That's a different thing entirely, and it's only possible with data at this scale."

One Architecture. Three Deployment Targets.

BADAS 2.0 ships as a family of three models. Each is optimized for a different deployment environment, all outperforming the prior generation on every metric:

• BADAS 2.0 (300M parameters): The large base model. State-of-the-art performance across all benchmarks. Best mean Time-to-Action and early warning recall. Expert in rare long-tail scenarios.
• BADAS 2.0 Flash (86M parameters): End-device-optimized. Engineered for false alarm prevention. Outperforms BADAS 1.0 on every metric.
• BADAS 2.0 Flash Lite (22M parameters): Ultra-light for IoT and edge deployment. Optimized for GPU and CPU inference. Rivals BADAS 1.0 performance at 14 times fewer parameters.

Flash Lite loses only 1 percentage point of Average Precision while running 12 times faster than the full model on A100 and 5 times faster on NVIDIA Thor. All three models operate within the 66-millisecond real-time budget at 16 Hz.

BADAS 2.0: Explainability, Reasoning, and Generalization

BADAS 2.0 extends the foundation with three capabilities that turn a warning into a response:

• Explainability. Attention heatmaps surface exactly what the model perceives and focuses on during risk scenarios, making incident prediction decisions auditable for regulated environments and enterprise deployments.
• Reasoning. The model predicts the appropriate response to a detected risk and explains its reasoning in natural language. BADAS 2.0 does not only predict an incident. It tells you what to do about it and why.
• Generalization. BADAS 2.0 demonstrates consistent incident prediction behavior across any physical collision scenario, including out-of-distribution, non-driving environments.

The Real World as Training Set

BADAS 2.0 was trained on the world's largest classified archive of naturalistic safety-critical events: 60 million edge-case videos drawn from 10 billion real-world miles. The breadth of that archive is visible in BADAS 2.0's performance across categories other models routinely fail: animals on road, fog, snow, and complex intersections.

Even fine-tuned on identical data, BADAS 2.0 outperforms a 2-billion-parameter foundation model across generalization benchmarks, including categories where the larger model fails most severely. The 22-million-parameter Flash Lite model alone outperforms the competition's full-scale offering. COSMOS achieves 77.7% on fog scenarios and 83.2% on infrastructure. All three BADAS 2.0 models exceed those figures.

BADAS 1.0 launched in October 2025 as the first video-native incident prediction model, trained on 400,000 real-world dashcam clips and deployed in commercial fleets beginning in December 2025. BADAS 2.0 was trained on 2,000,000. That scale is a structural condition of operating 350,000 cameras across 94% of US roads, capturing 100 million fresh miles every month, not a resource acquired for a single model release.

BADAS 2.0 is available now.

About Nexar

Nexar is the real-world intelligence platform for the Physical AI era — the independent verification infrastructure behind every machine that claims to know the road. A network of 350,000 cameras captures 100 million miles of real driving every month, producing the world's largest classified archive of naturalistic safety-critical events: 60 million edge-case videos, 10 billion miles of ground truth, 45 petabytes of verified road intelligence covering 94% of US roads.

BADAS 2.0, Nexar's incident prediction model family, achieves 99.4% Average Precision — #1 on all four major benchmarks. The platform serves Waymo, Lyft, IBM, NVIDIA, and government infrastructure clients simultaneously because Nexar competes with none of them. Independence is not a differentiator. It is the product.

Verifying AI. ✔ nexar.ai

Logo: https://mma.prnewswire.com/media/2509048/5903237/Nexar_Logo.jpg

SOURCE Nexar