Nvidia releases Alpamayo 2 Super for commercial robotaxi development

Nvidia released Alpamayo 2 Super, a 34 billion parameter vision language action model for commercial robotaxi development, available under the Linux Foundation's OpenMDW 1.1 license. The model, which can generate driving trajectories and reasoning traces from surround camera inputs, ranked first on the LingoQA benchmark with a Lingo Judge score of 79.2, outperforming Qwen2.5 VL 72B by 17 points, Gemini 2.5 Pro by 15.1 points, and GPT 4o by 23.2 points.

Nvidia releases Alpamayo 2 Super for commercial robotaxi development The 34 billion parameter model can generate driving trajectories, reasoning traces and training labels using full surround camera inputs. Nvidia released https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/ Alpamayo 2 Super for commercial use Tuesday, giving automakers and autonomous vehicle developers access to its largest open reasoning model for robotaxis and self driving systems. The model is available through Hugging Face under the Linux Foundation’s OpenMDW 1.1 license, which allows fine tuning, derivative models and commercial redistribution. Nvidia said it is also applying the license across earlier models in the Alpamayo family, which were initially released for research and development. Alpamayo 2 Super is a 34 billion parameter vision language action model comprising a 32 billion parameter Cosmos 3 Super Reasoner backbone and a roughly 2.3 billion parameter diffusion based action component. The model was post trained using reinforcement learning and is approximately three times the size of Nvidia’s 10 billion parameter Alpamayo 1 and Alpamayo 1.5 models. The model processes video from cameras positioned around a vehicle, alongside navigation instructions and its recent movement history. Nvidia said the full surround setup gives the system access to front, side and rear views when assessing situations such as lane changes, merges, intersections and unprotected turns. For each driving scenario, Alpamayo 2 Super can generate a planned trajectory and a chain of causation trace explaining the factors behind its decision. It can also produce higher level actions such as yielding, changing lanes or stopping. Additional functions include visual question answering, linking responses to regions within camera images and automatically generating reasoning labels for training and validation data. Nvidia said the model can be applied to proprietary fleet footage to convert raw driving clips into labeled training material. The Hugging Face model card says Alpamayo 2 Super was trained using approximately 115,000 hours of multi camera driving video and about 3.7 million chain of causation reasoning traces. Its outputs include text and planned vehicle trajectories covering up to 6.4 seconds. Nvidia said Alpamayo 2 Super ranked first among nearly 40 models tested on LingoQA, a benchmark focused on reasoning about autonomous driving scenarios. Using the Lingo Judge metric, Nvidia reported that the model scored 17 points higher than Qwen2.5 VL 72B, 15.1 points higher than Gemini 2.5 Pro and 23.2 points higher than GPT 4o. The results were reported from Nvidia’s own testing. The model card reports a Lingo Judge score of 79.2, alongside results from Nvidia’s closed loop AlpaSim evaluation and an open loop trajectory assessment based on challenging driving samples. Nvidia said developers can use Alpamayo 2 Super as a cloud based teacher model to generate data and reasoning outputs for smaller models. Those models can then be optimized for real time processing inside production vehicles. The broader Alpamayo platform also includes AlpaSim for closed loop simulation, AlpaGym for reinforcement learning, physical AI driving datasets and open training tools. Nvidia said Alpamayo models have collectively surpassed 500,000 downloads on Hugging Face. Disclosure: This article was edited by Estefano Gomez. For more information on how we create and review content, see our Editorial Policy https://cryptobriefing.com/editorial-policy/ .