benmagnifico / DriveJev

DriveJev 
 An open-source, efficient System I decision model for autonomous driving 
 Model  | 
 Live demo  | 
 Evaluation 
 DriveJev 1.0 at the wheel in JevPilot with full traffic. Top: an oncoming car turning left across our path, a 4-way stop shared with three cars, a red-light runner just after our green. Bottom: a pedestrian stepping out from behind a parked car, pedestrians at the exit of a left turn, three cut-ins on the interstate. Every drive reached its destination with no collision and no violation. Click a clip for the 720p video. 
 Three test situations: the front camera with the tele view (inset) and the probability DriveJev 1.0 gives each offered behaviour. 
> [!IMPORTANT]
> **News**
> - **[2026/10]** **DriveJev 1.0** is open source: the code, the model on Hugging Face (`benmagnifico/DriveJev-4B`), the JevPilot closed-loop harness and a live demo. On held-out closed-loop tests it drives 28 of 35 interaction episodes and 25 of 28 base episodes cleanly, meets 241 of 242 scripted hazards without contact and makes a decision in 111 ms on one GPU.
## Introduction
DriveJev is a System I driving model: like a fast, intuitive human driver, it does not write out its reasoning or plan a trajectory. At every decision it looks at three camera frames and a short description of the situation, scores the behaviours the car can carry out at that instant (*cruise*, *stop at the line*, *hold*, *start*, *turn*, *yield*, *brake*) and returns a probability for each, all in one forward pass. A local executor turns the chosen behaviour into steering and speed. DriveJev drives closed loop in the JevPilot town, city and interstate with traffic, pedestrians, traffic lights, stop signs and scripted multi-agent interactions.
Our contributions:
- **Driving ability.** DriveJev follows the traffic rules and drives steadily on city streets, small-town roads and the interstate, through junctions with traffic lights and with stop signs, among other cars and pedestrians. On held-out test drives it reaches the destination with no collision and no violation in 28 of 35 interaction episodes and 25 of 28 base episodes, and meets 241 of 242 scripted hazards (oncoming platoons, red-light runners, hidden pedestrians, cut-ins, 4-way-stop contention) without contact. In the same worlds, JevPilot's own autopilot, which asks the cloud Jev model to pick one of many sampled steering paths, does not hold a straight line in complex traffic.
- **Inference speed.** A decision is a single forward pass with no text or trajectory decoding. DriveJev decides in 111 ms (9 Hz) on one RTX 5090, while the Qwen-Drive-1.0 model it builds on runs at 1.2 to 1.7 Hz as a trajectory planner on the same GPU, and the cloud Jev API answers at about 4 to 5 Hz.
- **Extensibility.** Code, weights, simulator harness and demo are open and run locally on one GPU. DriveJev can be inspected, retrained and fine-tuned on new scenarios, which gives it room to grow towards complex and high-risk situations that a