The part worth writing about is not the architecture. It is what the architecture produced without being asked.
The model was trained on yes or no, and learned where to look
BADAS was trained on binary labels. Each clip was marked as containing a collision or not containing one. No bounding boxes, no annotated hazards, no human drawing a rectangle around the cyclist.
When the trained model runs, you can pull heat maps and bounding boxes straight out of its own attention on the forward pass, with no additional computation layered on top. It shows you the part of the scene it is reacting to.
That capability was never a training objective. It emerged. The team did not build an explainability feature; they discovered they had one.
Only we know
An emergent property is only observable to whoever trained the model and then went looking. This one came from binary labels alone, with no bounding-box supervision anywhere in the pipeline. A model that learns where to look from nothing but yes-or-no outcomes is telling you something about how much structure is already latent in real driving footage, and that finding is not available from a public dataset or a benchmark table. You have to own the library and the training run.
How it learns
The architecture is attention-based, trained in a V-JEPA style on latent space rather than on pixels. In practice that means the model encodes video into vectors, has parts of those vectors masked, and learns by predicting what was hidden. It is never asked to reconstruct an image. It is asked to complete a representation, which is a cheaper and more general thing to learn.
One trained model then drives several jobs through different heads: collision risk, video embeddings for search and clustering, and trajectory prediction.
Input is a two-second video clip. That is the observation window, and it is worth being precise about, because two seconds of input is a different quantity from any statement about warning time. This piece makes no claim about the second thing.
Why 200,000 videos and not 10 billion miles
The library behind BADAS holds more than 10 billion miles of fleet driving and more than 70 million long-tail driving videos, where long tail means the rare road situations that almost never happen and matter most. The trained model saw 2 million clips drawn from 200,000 videos.
The gap between those numbers is the whole difficulty of the problem. Negative examples, meaning ordinary driving where nothing happens, are effectively unlimited and repeat themselves, so there is little value in training on all of them. Positives are hard to get. A collision library is not built by collecting more driving. It is built by finding the fraction of driving that contains an event.
The method for finding it is a loop. The previous version of the model scores the entire video archive, and the clips it finds suspicious go to human annotators who confirm whether a collision is present. The model mines its own successor's training set. That is also how false positives surface: borderline clips that look like they are about to become a collision and do not.
That produced 5 times the training data of the previous version.
Three sizes, and where each one can actually run
BADAS ships as a 300 million parameter Large, an 86 million Flash, and a 22 million Flashlight. The two smaller models are distilled from the Large.
Distillation here is doing less work than the name suggests. Most of the model's understanding comes from the self-supervised pre-training; the distillation step mainly adjusts where the model draws its decision boundaries. The small models are not lesser models that learned less. They are the same understanding, tuned to fit hardware.
What it does not do
The written explanation of a hazard does not come from BADAS. It comes from a separate head, a fine-tuned Qwen 3 vision-language model trained on frames with bounding boxes plus reasoning and action text. Two different mechanisms, and conflating them is the easiest mistake to make about this system. The attention heat maps are the model showing where it looked. The sentence is a second model describing a scene.
The reasoning head is too heavy to run on the edge. Only the attention maps can run in real time there. Anyone reading a natural-language hazard explanation as an in-vehicle feature has the architecture wrong.
The result the team did not expect
BADAS ran on customer data without any fine-tuning. The expectation was that customer-specific tuning would be needed, because that is how these deployments normally go. It was not needed, and that repeated across different customers.
No customer is named here and no count is stated, because neither is settled evidence.
Limits
The figures above describe one trained release. The library size is Nexar's fleet library, not a combined company figure. Model architectures move, and a parameter count is a fact about a specific checkpoint rather than about a product line.
This piece deliberately makes no claim about how far ahead of an event the model raises risk, and no claim about accuracy against any benchmark. Both exist. Both are in review.
Measured, not imagined. The model learned where to look because the footage was real.