NVIDIA Open-Sources Alpamayo 2 Edge, a 2B Self-Driving Reasoning Model That Runs In-Car
On this page 9 sections
TL;DR
Alpamayo 2 Edge is NVIDIA's new open-weights driving model: a 2B-parameter reasoning vision-language-action network that watches six cameras, plans a 64-waypoint trajectory and explains its own decision, quickly enough to run inside a DRIVE AGX Thor car computer. With the deployment recipe NVIDIA published Nov 10, pixels become trajectory in 160-190 ms. Weights ship under OpenMDW-1.1, code is Apache 2.0, and a reasoning judge, an RL gym, a simulator and two datasets come alongside.
What shipped, exactly
The weights sit on Hugging Face in BF16, directly runnable as a PyTorch model. The source repo is Apache 2.0 and carries the inference script, a fine-tuning path and a visual question-answering notebook. Around the model NVIDIA published the rest of the toolchain: the AlpaJudge reasoning judge, the AlpaGym closed-loop RL environment, the AlpaSim simulator that scores the driving, and the PhysicalAI-AV dataset with a NuRec reconstruction variant for closed-loop tests.
The technical blog comes from Marco Pavone, NVIDIA's VP of autonomous systems research and a Stanford professor, and Boris Ivanovic, who manages the group. Alpamayo itself already has takers: Bolt picked it for the robotaxis it plans to launch in Madrid in 2027, and Einride is putting it in autonomous trucks.
Two heads: one thinks, one drives
The backbone is the Nemotron reasoning core of Cosmos 3 Edge: 2.45B parameters, split as 1.96B language model, 0.41B SigLIP2 vision encoder and 0.08B projector. Bolted to it sits a 0.2B diffusion action expert, and the pair is what NVIDIA calls a vision-language-action model.
Each observation is 22 frames across six cameras: five frames each from the front-wide and front-tele cameras, three each from the cross and rear pairs, plus navigation text and 20 ego poses at 0.1 s steps. The output is 64 waypoints covering 6.4 s of driving. The action expert predicts per-step acceleration and curvature, and a unicycle model integrates them into the path.
Distilled from Chain-of-Causation autolabels generated by the bigger Alpamayo 2 Super, the model narrates only when the road asks for it. In routine driving it emits a <NOMINAL> token instead of an explanation, which keeps the token budget for the moments that matter. A model that knows when to stop narrating is ahead of most group chats.
Then came closed-loop RL in AlpaGym, where the model learns from the consequences of its own actions in simulation. NVIDIA credits that stage with a 48.9% improvement in closed-loop driving over the best checkpoint without it.
The scoreboard, and who keeps it
NVIDIA's comparison point is Qwen-Drive-1.0, a 4B vision-language-action driving model, so the smaller model beat a bigger one on the vendor's own track. In AlpaSim, over 875 NuRec scenarios with three rollouts each, Alpamayo 2 Edge logged 15.15 km between at-fault incidents, where at-fault means off-road or at-fault collision, against 0.25 km for Qwen-Drive-1.0. That is the 60x headline.
Open-loop, the blog's chart puts path error at 0.53 m against 1.68 m for Qwen-Drive-1.0, and the model card logs minADE_6 at 5 s of 0.52 m across 1,392 PhysicalAI-AV samples. Scene understanding runs on the usual driving-VQA boards, from LingoQA to WaymoQA:
NVIDIA built the simulator, picked the opponent and reported the gap, so treat the 60x as a vendor claim until someone independent reproduces it. Everything above is simulation or open-loop logs.
Squeezing it onto the car
On a workstation the BF16 model runs in plain PyTorch on any supported NVIDIA GPU. The card's measured configuration, one 22-image sample on an H100 80GB, peaks around 12 GiB, and NVIDIA recommends 16 GB of GPU memory. Real-time on Thor is another job: the Developer Hub recipe walks through FP8 scale calibration, ONNX export and TensorRT engines with Edge-LLM plugins, plus DFlash2 speculative decoding.
Speculative decoding pairs a junior with a senior: the small draft model blurts a guess at the next tokens, the big model confirms or corrects, and checking is cheaper than writing from scratch. On DRIVE AGX Thor-X, under DRIVE OS 7.2.7.0 with CUDA 13.4 and TensorRT 11.3.1.9, the finished stack runs pixel-to-trajectory in 160-190 ms depending on how much it reasons.
Early partners report fast uptake. Embedl had it planning at 10 Hz on AGX Thor in under 48 hours. PORT, testing on reconstructions of its own London street recordings, kept its lane in twice as many runs as Alpamayo 1.5 and found the out-of-the-box model planned roughly twice as fast as its hand-tuned 1.5 server on the same GPU. Niulinx measured 13.6% lower displacement error and 46.1% lower lateral error at 3 s than Alpamayo 2 Super in a zero-shot test, then cut displacement error another 44.2% by fine-tuning on its own driving data, with a real vehicle running within days.
AlpaJudge grades the explanations
Shrinking a reasoning model risks fluent nonsense, so NVIDIA also released a judge. AlpaJudge is human-calibrated and comes in two variants. RB scores a candidate explanation against a reference, text only. RF looks at six camera frames plus the car's action history and returns five rubric probabilities, from action agreement to causal consistency.
The blog shows two uses. For data selection, resampling toward the mid-score clips a model still learns from raised their share from 27% of the raw set to 81% of the curated set. As an RL reward, on an Alpamayo 2 Super-based model, the judge's score lifted reasoning quality 10.2% by its own metric and 11.4% by human evaluation. A public rubric judge feeding both curation and rewards is a loop any agent team can copy.
The catches
- Every benchmark in this post is NVIDIA-run. The closed-loop numbers come from the company's own AlpaSim against a baseline NVIDIA chose, and no third party has published results.
- The 60x figure is a simulation result over 875 scenarios, with no road testing behind it.
- Real-time in-car deployment assumes DRIVE AGX Thor hardware. On the PyTorch path, other GPU architectures beyond H100 are explicitly not validated.
- The datasets sit behind access terms on Hugging Face, so this is not a no-questions download.
- The model card's own warning applies to any AV use: integration requires use-case-specific testing, and an explanation that reads well is debugging evidence, not a safety case.
Why builders should care
If you work on autonomy, this is a foundation model you can own. Fine-tune it on your fleet's data the way Niulinx did, evaluate it in your own scenarios, and keep the whole loop in-house under an open license.
If you build agents outside cars, the transferable part is the loop, and every piece of it is public: distill reasoning from a big teacher, post-train in a closed-loop gym, grade explanations with a rubric judge, curate data by judge score, then feed the same judge back in as the reward. Swap driving for your domain and the recipe still reads.
If you ship models onto tight hardware, the Thor recipe is a working reference for the full compression stack: FP8 calibration, ONNX export, TensorRT engines and speculative decoding, with per-stage measurements you can reproduce.
Key Takeaways
- NVIDIA released Alpamayo 2 Edge on Nov 10: a 2B-parameter reasoning VLA driving model, 2.45B backbone plus 0.2B diffusion action expert, under OpenMDW-1.1 weights and Apache 2.0 code, with AlpaJudge, AlpaGym, AlpaSim, deployment recipes and the PhysicalAI-AV datasets published alongside.
- It takes 22 frames from six cameras plus navigation text and ego history, and outputs 64 waypoints over 6.4 s; routine driving emits a NOMINAL token instead of reasoning, saving tokens for situations that need them.
- On DRIVE AGX Thor-X the TensorRT-plus-DFlash2 build runs pixel-to-trajectory in 160-190 ms, and Embedl reached 10 Hz planning on AGX Thor in under 48 hours.
- NVIDIA reports 15.15 km between at-fault incidents in AlpaSim against 0.25 km for the 4B Qwen-Drive-1.0, and open-loop path error near 0.52-0.53 m; all results are vendor-run, in simulation or open-loop logs.
- Closed-loop RL in AlpaGym added 48.9% to driving performance; AlpaJudge-based curation raised informative mid-score clips from 27% to 81% of the training set, and judge rewards lifted reasoning quality 10.2-11.4%.
- Partners report portability: Niulinx cut displacement error 13.6% zero-shot against Alpamayo 2 Super and another 44.2% after fine-tuning on its own data, reaching a real vehicle within days.
Sources: NVIDIA Technical Blog, Nov 10, 2026, Alpamayo 2 Edge model card, Hugging Face, NVlabs/alpamayo2-edge, Alpamayo Developer Hub recipes, AlpaJudge, Hugging Face, AlpaGym, AlpaSim, PhysicalAI-AV dataset, LingoQA, EVwire, Nov 11, 2026