NVIDIA Alpamayo 2 Super is a 34B open vision-language-action model for robotaxis. Released under OpenMDW-1.1, it tops LingoQA with 79.2 and excels at long-tail driving.
NVIDIA has released Alpamayo 2 Super, a 34B open vision-language-action (VLA) model purpose-built for autonomous driving and robotaxi development. Unlike modular perception stacks that separate detection, prediction, and planning into discrete subsystems, this end-to-end model generates a planned trajectory, a natural-language Chain-of-Causation (CoC) explanation, and a meta-action label in a single forward pass over full-surround camera video. Released under the Linux Foundation’s permissive OpenMDW-1.1 license, with source code under Apache 2.0, Alpamayo 2 Super is immediately available for commercial use, fine-tuning, and redistribution. For AV teams wrestling with the long tail of edge cases—unusual intersections, unpredictable pedestrians, dense multi-agent traffic—this is a significant step forward.
The model pairs NVIDIA’s 32B Cosmos 3 Super Reasoner vision-language backbone with a 2.3B diffusion-based action decoder. The full model totals 34B parameters, a substantial yet manageable footprint. In BF16 precision, it runs on a single H100 80GB GPU with a peak memory of 72,115 MiB.
The design goal is to handle long-tail, multi-agent driving scenarios that traditional detection-and-prediction pipelines struggle to process. Rather than treating perception, prediction, and planning as separate stages with hand-crafted interfaces, Alpamayo 2 Super learns directly from video and action data. This holistic mapping from visual context to driving behavior often proves more robust in messy, real-world conditions.
From six cameras with four historical frames each, the model produces:
This explainable-action design is a key selling point. Engineers and safety reviewers don’t have to trust a black box. They can inspect the CoC trace to understand why the model decided to brake, accelerate, or change lanes—an essential capability for validation, debugging, and regulatory oversight.
NVIDIA evaluated Alpamayo 2 Super on LingoQA, a standard benchmark for vision-language understanding in driving, and reports a Lingo-Judge score of 79.2. Among nearly 40 evaluated models, it ranked first. The margins over larger general-purpose VLMs are striking:
In short, a purpose-built 34B driving model beats generalist models two to four times its size. Specialization matters, and the numbers show it.
The model also performs well on trajectory planning benchmarks. On AlpaSim closed-loop planning, it scored 1.50 ± 0.13 across 910 scenarios from the PhysicalAI-AV-NuRec dataset. In open-loop evaluation, it posted a minADE₆ of 0.911 meters at 6.4 seconds on 937 challenging samples from the PhysicalAI-AV dataset.
For context, lower minADE₆ values indicate more accurate trajectory prediction. Sub-meter error at 6.4 seconds on challenging samples suggests the model can anticipate complex road geometries and vehicle interactions effectively—critical for robotaxi safety and passenger comfort.
The model’s capabilities come from a substantial, driving-specific training corpus:
The CoC traces are especially valuable. They pair natural-language explanations with driving actions, teaching the model not just what roads look like but how to reason about them and act accordingly. This is what enables grounded visual question answering and the explainable outputs that make the model useful for safety validation workflows.
One of the most significant aspects of Alpamayo 2 Super is its licensing. Released under the Linux Foundation’s OpenMDW-1.1 open-model license, with source code under Apache 2.0, it offers:
This is a marked shift from the “open weights but restricted use” approach common elsewhere in the AI industry. For AV companies, it removes a major legal hurdle to adopting open models in production.
NVIDIA positions Alpamayo 2 Super as an effective autolabeler. Proprietary fleet data that might take months to annotate manually can be processed in days. The model generates reasoning auto-labels—explanations attached to driving data—making dataset curation faster, cheaper, and more consistent across teams.
The model also supports grounded visual question answering. Developers can query a scene, ask “Is the vehicle ahead slowing down?”, and receive answers tied to specific observations. This fits naturally into NVIDIA Halos safety-validation workflows, helping teams connect observations to actions and surface failure modes before on-road deployment.
Alpamayo 2 Super is designed as a cloud-scale model. NVIDIA advises distillation for in-car inference: use the full model as a teacher to train smaller student models that run in real time on embedded hardware. This teacher-student pipeline is standard practice in AV development, where on-vehicle compute is constrained but accuracy requirements remain high.
The validated public notebook used six cameras with four historical frames per camera and peaked at 72,115 MiB on an H100 80GB in BF16. That’s a serious but workable footprint for a cloud-side development flow, and a clear sign that an in-car version will arrive via distillation.
Alpamayo 2 Super sits at the intersection of several converging trends:
This wave of open, production-ready models could accelerate the AV industry by making state-of-the-art research directly available to startups, OEMs, and mobility providers. The combination of permissive licensing, explainable outputs, and strong benchmarks lowers the barrier to entry for teams that previously had to build everything from scratch.
NVIDIA Alpamayo 2 Super marks a milestone in open autonomous driving AI. Its 34B parameter architecture, top-tier LingoQA performance, and permissive OpenMDW-1.1 license make it an attractive foundation for robotaxi and ADAS development. The model’s explainable Chain-of-Causation outputs, autolabeling capabilities, and strong trajectory planning results offer both immediate value and a clear path to production.
For teams building robotaxi systems, ADAS features, or AV simulation pipelines, the next step is straightforward: download the model, fine-tune it on proprietary fleet data, and validate it within your safety framework. The open era of autonomous driving software has arrived—and it’s accelerating fast.
Alpamayo 2 Super is a 34-billion-parameter open vision-language-action (VLA) model for autonomous driving and robotaxis. It combines a 32B vision-language backbone with a 2.3B action decoder to generate a planned trajectory, a Chain-of-Causation explanation, and a meta-action label in one forward pass.
Traditional modular stacks separate detection, prediction, and planning into discrete subsystems with hand-crafted interfaces. Alpamayo 2 Super instead learns an end-to-end mapping directly from full-surround camera video to driving actions, which can be more robust in long-tail, multi-agent scenarios like unusual intersections or unpredictable pedestrians.
Yes. The model is released under the Linux Foundation's permissive OpenMDW-1.1 license, and its source code is under Apache 2.0. That means it is immediately available for commercial use, fine-tuning, and redistribution, making it accessible for AV teams and robotaxi developers.
In BF16 precision, the model runs on a single NVIDIA H100 80GB GPU with a peak memory of 72,115 MiB. The full model has 34 billion parameters, so while it is a substantial model, it remains manageable on enterprise-grade hardware.
The Chain-of-Causation is a natural-language explanation the model produces alongside its planned trajectory, describing why it chose a particular action such as braking, merging, or yielding. This is important for safety review and debugging because it lets engineers inspect the model's reasoning instead of treating it as a black box.
