NVIDIA Releases Alpamayo 2 Edge Open Weights for Self-Driving Cars
NVIDIA has released the open weights for Alpamayo 2 Edge, a compact reasoning model built to drive a car from inside the vehicle itself. The release came on 10 November 2026, in a developer blog post by Marco Pavone and Boris Ivanovic titled "Running AV Reasoning Models Within Onboard Compute and Latency Limits", and NVIDIA announced it on X at 16:08 UTC the same day. A vision-language-action model, or VLA model, is a model that takes in camera images and text instructions, reasons about what it sees, and outputs driving actions such as steering and speed. Alpamayo 2 Edge is a 2-billion-parameter VLA model for autonomous driving, and its weights are now on Hugging Face under the OpenMDW-1.1 licence, with inference code on GitHub under Apache 2.0. This is a weights and code release, not the first announcement: Alpamayo 2 was first announced on 21 October 2026.
What Alpamayo 2 Edge is
Alpamayo 2 Edge is built on the Cosmos 3 Edge Nemotron backbone, a 2.45-billion-parameter model, plus a 0.2-billion-parameter diffusion action expert, which together come to about 2.65 billion parameters in components and are marketed as a 2-billion-parameter model. It takes in six camera feeds, navigation text, and the car's own motion data, and it outputs a driving trajectory: 64 waypoints covering the next 6.4 seconds. The model's signature feature is adaptive reasoning. In routine driving it emits a single "NOMINAL" token, a short signal that means the situation is ordinary and needs no extended thought. Only when it faces a complex hazard does it switch to full Chain-of-Causation reasoning, a step-by-step account of what is happening and what it should do about it.
Built for the car, not the cloud
Alpamayo 2 Edge is designed to run on NVIDIA's DRIVE AGX Thor, the in-vehicle computing platform, with a target of 10 hertz, meaning the model updates its driving decisions ten times a second. The optimised recipe runs at 160 to 190 milliseconds from pixel to trajectory on Thor-X, using DRIVE OS 7.2.7.0, CUDA 13.4, and TensorRT 11.3.1.9, with FP8 calibration, ONNX, TensorRT, and DFlash2 in the pipeline. There is an important qualification. The downloaded BF16 weights run as PyTorch on supported GPUs, but real-time operation on Thor needs further optimisation beyond the download, and the baseline compatibility was tested on a single H100 80GB card. In other words, the weights are open and runnable, but turning them into a real-time in-car system is still engineering work. The model card is nvidia/Alpamayo2-Edge on Hugging Face, and the code repository is NVlabs/alpamayo2-edge on GitHub.
AlpaJudge: the evaluator that comes with it
Released alongside the model is AlpaJudge, a human-calibrated evaluation and reward model. Its job is to judge the quality of the driving model's reasoning, to select the best training data, and to provide rewards for reinforcement learning. AlpaJudge is released under the same weight and code licences as Alpamayo 2 Edge. According to the 10 November digest, it works in two modes, reference-based and reference-free, but that split is stated in the digest only, so it is presented here as attributed detail rather than as a confirmed feature of the release. The broader point stands without it: NVIDIA is releasing not just a driving model but the tool it uses to judge and improve that model's reasoning.
 |
| An autonomous car driving on a highway at dusk with a roof-mounted sensor pod visible. |
The benchmarks: strong, and simulated
NVIDIA's benchmarks for Alpamayo 2 Edge are striking, and they need to be read with care. In the AlpaSim closed-loop simulator, the model drove 9.87 kilometres between at-fault incidents, compared with 0.25 kilometres for the 4-billion-parameter Qwen-Drive-1.0, a difference NVIDIA describes as about 40 times; the exact ratio is 39.5. Closed-loop reinforcement learning in AlpaGym improved driving performance by 48.9 percent compared with the best checkpoint trained without it. All of these numbers are NVIDIA's own, and all of them come from simulation. They are not road-safety records, and they should not be reported as proof that the model is safe to drive on public roads. A simulator measures performance in a controlled environment; it does not measure what happens when a real car meets a real road.
| Measure | Alpamayo 2 Edge | Comparison |
|---|---|---|
| Distance between at-fault incidents (AlpaSim) | 9.87 km | 0.25 km for 4B Qwen-Drive-1.0 (about 40x) |
| Distance per at-fault incident (model card, 875 scenarios) | 15.15 km | Separate figure, same model |
| Closed-loop RL gain (AlpaGym) | +48.9% driving performance | Versus best checkpoint without RL |
| Training data | About 115,000 hours of footage | Plus 3.7m Chain-of-Causation traces (digest figure) |
| Latency (optimised, Thor-X) | 160-190 ms pixel-to-trajectory | 10 Hz target |
Edge, Super, and the rest of the family
Alpamayo 2 Edge should not be confused with the other members of NVIDIA's AI family. The 34-billion-parameter Alpamayo 2 Super is a separate, larger model built for the cloud or as a teacher model, and the two should not be merged in reporting. Alpamayo 2 Edge is the compact in-vehicle version, the one that runs on Thor inside the car. It is also distinct from GR00T 2, NVIDIA's model for humanoid robots, and from Cosmos 3 Nano, a smaller model in the Cosmos family. GR00T 2 moves a robot's body, Cosmos 3 Nano generates video, and Alpamayo 2 Edge drives a car. Keeping the names straight matters, because each model has its own benchmarks, its own licence, and its own hardware.
Why open weights matter for self-driving
Releasing open weights for a driving model is a different kind of decision from releasing a chatbot. Driving models have historically been closed, held inside the companies that build robotaxis and driver-assistance systems, because the data and the safety questions are so sensitive. NVIDIA's release gives researchers, universities, and smaller companies a working reasoning model for driving that they can download, inspect, and build on, under an open licence. Our coverage of the Mistral Large 4 release and the Kimi K3.1 Pro launch shows how open weights have become a competitive strategy in their own right: the model is free, and the value lies in the hardware and the platform it runs on. Our flagship AI model comparison shows how the big language models are judged on reasoning, coding, and knowledge; Alpamayo 2 Edge is judged instead on whether a car stays in its lane and avoids a collision. For NVIDIA, that logic is direct. Alpamayo 2 Edge runs best on Thor, so giving away the model can sell the chip. For the research community, the value is a real, working VLA model with a published evaluation tool, which is rarer than it should be in autonomous driving.
 |
| A car dashboard with a driver-assist display showing the road ahead and lane markings. |
What the adaptive reasoning actually buys
The adaptive reasoning design deserves a closer look, because it is the model's main technical claim. Most reasoning models think at full depth on every input, which is expensive and slow. Alpamayo 2 Edge does the opposite: it treats routine driving as a case for a single token, and reserves full reasoning for hazards. That is a bet that most driving is boring, and that the model's job is to notice the moments that are not. The 9.87-kilometre figure in the simulator is NVIDIA's evidence that the bet works in the simulated world, but the real test is whether the classification holds up in the messy, unscripted situations that simulators struggle to reproduce. That is why the release of the weights matters more than the benchmark: researchers can now test the claim themselves, rather than taking NVIDIA's word for it.
 |
| Close-up of a car's front sensor and camera array integrated into the grille and bumper. |
Key Takeaways
- NVIDIA released the open weights for Alpamayo 2 Edge on 10 November 2026; the model was first announced on 21 October, so this is a weights and code release, not a first announcement.
- It is a 2-billion-parameter reasoning vision-language-action model that takes six camera feeds, navigation text, and ego-motion, and outputs 64 waypoints over 6.4 seconds.
- Adaptive reasoning uses a single "NOMINAL" token for routine driving and full Chain-of-Causation reasoning only for complex hazards, targeting 10 Hz on DRIVE AGX Thor.
- AlpaJudge, a human-calibrated evaluation and reward model, was released alongside it under the same licences.
- NVIDIA's own simulation benchmarks report 9.87 km between at-fault incidents versus 0.25 km for Qwen-Drive-1.0, and a 48.9 percent gain from closed-loop reinforcement learning — simulation figures, not road-safety records.
Conclusion
Alpamayo 2 Edge is a serious and useful release: a compact reasoning model for driving, with open weights, open code, and an evaluation model to go with it, built to run inside a car on NVIDIA's Thor platform. Its adaptive reasoning design is a credible answer to the problem of running a thinking model at driving speed, and its simulation benchmarks are strong. But the benchmarks are simulations, and the real test will be independent researchers running the open weights and checking whether the model's judgement holds up outside NVIDIA's own simulator. That testing can now begin, which is the most important thing about the release.
Comments
Post a Comment