AREN · technical overview
A 3B-parameter multimodal model for the car.
AREN is a 3-billion-parameter multimodal model for automotive intelligence, designed to run on production automotive SoCs. This page is the current, plain reading of what AREN is and what it is not.
Model card
Current design targets. Unknown rows are left as em dashes.
| Property | Value |
|---|---|
| Parameters | 3B |
| Architecture | Multimodal transformerdetail on request under NDA |
| Input modalities | Camera, LiDAR, radar, CAN bus, natural language |
| Output | Perception, planning, low-level control hints |
| Precision formats | FP16, INT8, INT4target |
| Target SoCs | NVIDIA DRIVE Thor, Snapdragon Ride Elite, Exynos Auto-class |
| Target OS environments | QNX OS for Safety, VxWorks, Linux with PREEMPT_RT |
| Latency target | — |
| Power envelope | — |
| Status | In active development |
Current AREN design targets. Update cadence: research notes.
Design principles
One model, not a pipeline.
The classical stack — detection, tracking, prediction, planning, control — was structurally right when compute was scarce and datasets were small. Both constraints have lifted. Each module hand-off in the cascade throws away information the next module would benefit from, and the long tail of driving lives at the seams. A single learned model keeps the representation intact end to end.
Compactness as a safety property.
AREN is 3B parameters, not 30B, because the vehicle budget is real: thermal envelope, memory, deterministic latency. A model that fits the SoC of a shipping vehicle is the model that gets deployed. A model sized for a datacenter never leaves the datacenter.
Training as a first-class system.
Data curation, simulation, and curriculum design are not adjunct to the model — they are the model. Our research notes on AREN are as much about the training system as the weights.
Safety posture
Hardware integration
AREN targets automotive NPU or GPU partitions on NVIDIA DRIVE Thor, Qualcomm Snapdragon Ride Elite, and Samsung Exynos Auto-class SoCs. These are technical integration targets, not customer claims.
Deployment assumes a certified hypervisor partitioning AREN from the safety island. Inference is mixed-precision; INT8 is our working target, INT4 is under evaluation for specific subgraphs. Concrete latency and power figures are left as em dashes on the model card until measured in a reference vehicle configuration.
Serious integration inquiries route to Contact with inquiry type “Partnership.”