Nvidia Open-Sources Alpamayo 2 Super Reasoning Model for AVs

When a chipmaker whose margins depend on training compute open-sources a frontier model, it is worth reading twice. NVIDIA has done exactly that with Alpamayo 2 Super, a reasoning model aimed at robotaxis and autonomous vehicles, released under the Linux Foundation's OpenMDW-1.1 license that permits fine-tuning, derivative models, and commercial redistribution. The company announced it on August 4 on its corporate blog.

The interesting design choice is what the model outputs. It does not just emit a trajectory. It produces five coupled outputs, including chain-of-causation reasoning traces, meta-actions like yield and lane changes, and visual question answering with 2D image grounding across full-surround camera coverage. Those causation traces, according to NVIDIA, plug into its Halos safety-validation workflows and support AI safety aligned with ISO/PAS 8800 requirements. In other words, the model is being pitched as auditable, not just accurate.

On benchmarks, NVIDIA's claim is that Alpamayo 2 Super ranks first on LingoQA, an autonomous driving reasoning benchmark, among nearly 40 models evaluated, beating Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1, and GPT-4o by 23.2. Built on the Cosmos 3 Super Reasoner and post-trained with reinforcement learning, it is 3x the scale of the 10-billion-parameter Alpamayo 1.5 and Alpamayo 1 models, and the Alpamayo family has already surpassed 500,000 downloads on Hugging Face.

The honest caveat is that these are vendor-run numbers on a driving-QA task, not on real long-tail safety, and the post does not name a single OEM or robotaxi operator already validating this specific model in production. What the reporting also does not give you is the exact parameter count, the inference hardware target, or the latency budget for on-vehicle use. Take the LingoQA scores as a directional signal, not settled ground truth.

Still, the direction matters. A permissively licensed reasoning model with inspectable causation traces changes the equation for AV startups whose differentiation used to be model quality. The moat is quietly shifting toward data pipelines, integration, and the safety validation stack, which, conveniently for NVIDIA, is exactly the layer where Cosmos and Halos live.