TIER IV and Autoware Release AutoE2E Driving Model

TIER IV and the Autoware Foundation on 11 November 2026 released AutoE2E, an open-source end-to-end driving model with 79.9 million parameters, as part of a wider Autoware 2.0 model drop. In a TIER IV press release and an Autoware Foundation post the same day, the two organisations said the Reactive model is out now, while a World Model and a Reasoning Model are still in development.

The Hugging Face model card matches the 79.9 million-parameter figure and is the document that sets the legal terms for the weights. The AutoE2E source code on GitHub is under Apache 2.0. The v1.0 checkpoints are under CC BY-NC-SA 4.0. The card is plain: “commercial use is not covered.” It needs permission from upstream rights holders, including a commercial licence from Motional for nuPlan and nuScenes. The Foundation post’s line that the code is available “for research and commercial use” refers to the Apache 2.0 code, not to those checkpoints, and the two licences should not be blurred.

TIER IV also announced METEOR, a 16 million-parameter model that it says runs on NVIDIA and Renesas hardware, plus OnePlanner, LatentDrive, and an integration of NVIDIA’s 2 billion-parameter Alpamayo 2 Edge with Autoware. AutoE2E is the lead of that group, and the only model this post treats in detail.

AutoE2E’s published score is an open-loop number on the Foundation’s own benchmark

On KITScenes-Multimodal validation data, the Foundation says a fine-tuned AutoE2E checkpoint “achieved an average displacement error (ADE) of 0.744 meters over a three-second horizon, ranking second among eight open-weight model configurations.” The card repeats that ranking and stresses the method: each model predicts the next three seconds from logged camera frames and is scored against the path a human driver took. That is open-loop evaluation, not a closed-loop test of the car in traffic. The card also notes that AutoE2E receives a map raster and route in that comparison, and that inputs differ by model.

Real-vehicle closed-loop testing “is underway,” the Foundation says. “Further on-road trials are planned for 2027.” L4 capability “remains a development goal.” It is not a certification, and the model card adds that AutoE2E v1.0 “is not a certified driving system” and is not for controlling a vehicle on public roads. The card’s intended uses are open-loop evaluation, visualisation, benchmarking and further training.

Architecturally, AutoE2E is described as three cooperating models: a Reactive planner running at 10 Hz, a world model at 1 Hz, and a reasoning model at 1 Hz. Only the Reactive layer is in the v1.0 checkpoints. The other two branches are “disabled in these weights and remain under active development.” Every published number above, the Foundation says, was achieved by the Reactive model alone.

An open driving model is not a licence to put it in a paid robotaxi

The licence split is the part of this release that will trip people who only read the Foundation’s “research and commercial use” line. The code can be used commercially under Apache 2.0. The weights cannot, at least not under the terms the Autoware Foundation has given itself, because they were trained on nuPlan, nuScenes and KITScenes data whose own terms are non-commercial. Anyone hoping to fine-tune AutoE2E into a product still has to obtain those upstream rights.

The rest of the announcement should be read in the same cautious register as other 2026 driving-stack news. Bolt has already said it plans European robotaxis on NVIDIA Alpamayo, starting in Madrid. Volvo and NVIDIA have expanded a door-to-door automated-driving pact that also sits on the Alpamayo family. Those are company plans, and both of those TSN posts remain in-text references only. Closer comparisons that are still being built out include Mercedes-Benz’s extended NVIDIA partnership on automated driving and Freenow and Baidu’s RT6 robotaxi plans for London, with a 2027 launch aim. AutoE2E’s own 2027 item is more modest: further on-road trials, not a certified L4 service.

The honest summary is small and useful. A 79.9 million-parameter Reactive model is available, with an open-loop ADE of 0.744 metres over three seconds on the Foundation’s own eight-model comparison. The world and reasoning models are not out. The code is Apache 2.0; the checkpoints are not commercially licensed. L4 is a goal. The road tests are planned for 2027.