NVIDIA는 자율주행을 위한 상용 라이선스 기반 오픈 추론 모델인 Alpamayo 2 Super를 출시했으며, 이는

For robotaxis and other autonomous vehicles, the greatest challenges lie in rare, complex situations rather than everyday driving scenarios. Addressing these long-tail events requires systems that can understand contexts, reason about cause and effect, and execute safe decisions in real time. To solve this, NVIDIA has made Alpamayo 2 Super available for commercial use. As part of the Alpamayo family, which stands as the most-adopted open reasoning models for autonomous driving on Hugging Face, it consolidates a wide range of AV-relevant capabilities into a single foundation model built on NVIDIA Cosmos 3 Super Reasoner and refined through reinforcement learning.

The model is distributed under OpenMDW-1.1, the Linux Foundation’s permissive license for open AI model distributions. This license permits fine-tuning, derivative models, and commercial redistribution, allowing automakers, truckmakers, suppliers, and developers to adapt the system to their own data, policies, and deployment strategies. By applying this open licensing across the entire Alpamayo family, NVIDIA ensures teams retain full control over their proprietary data and infrastructure while owning the value they create. Open weights also make development economically viable by eliminating the need to rebuild foundational capabilities from scratch or pay premium costs for every task, enabling teams to match the right model to the right job at the right price.

Alpamayo 2 Super enables frontier-scale reasoning within cloud-based development workflows, where engineers generate high-quality reasoning traces, synthetic training data, and teacher outputs for model distillation. Within the family, Alpamayo 2 Super provides the highest reasoning and driving performance for multimodal development, while Alpamayo 1.5 and Alpamayo 1 offer more cost-efficient alternatives for cloud development and distillation. The resulting distilled models are optimized for efficient, real-time inference in production vehicles. This creates a complete cloud-to-car workflow that combines advanced cloud reasoning with scalable deployment across commercial fleets, offering a sustainable path to scaling safe autonomy.

In independent evaluations, Alpamayo 2 Super ranks first on LingoQA, an autonomous driving reasoning benchmark tested against nearly forty models. During NVIDIA testing using the Lingo-Judge metric, it outperformed Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points, demonstrating state-of-the-art reasoning for driving-centric scenarios. It also holds the top position across all autonomous driving benchmarks evaluated by NVIDIA. The model features three times the scale of the 10-billion-parameter Alpamayo 1.5 and Alpamayo 1 models, granting it superior capacity to generalize reasoning from sparse examples, which is critical for navigating rare, multi-agent interactions that typically challenge conventional systems.

The system reasons over full-surround camera coverage, fusing views from the front, sides, and rear to provide 360-degree context that improves understanding of lane changes, merges, unprotected turns, and complex intersections. For each driving situation, it produces five tightly coupled outputs: a trajectory describing the planned path, a chain-of-causation trace explaining the decision logic, a meta-action capturing intent such as yielding or stopping, reasoning auto-labels that generate annotations for training and validation, and visual question answering responses with 2D visual grounding that link answers to specific image regions. These outputs give developers clear visibility into the decision-making process, making it easier to critique and validate actions. The chain-of-causation traces integrate directly with NVIDIA Halos safety-validation workflows and support artificial intelligence safety aligned with ISO/PAS 8800 requirements.

Beyond planning, the model functions as an autolabeler that generates chain-of-causation labels and performs visual question answering on proprietary fleet data, compressing annotation cycles from months down to days. Its multitask design also supports scene understanding, model critiquing, and knowledge distillation, allowing developers to rely on a single foundation model across much of the development stack to simplify tooling and speed up iteration. Alpamayo 2 Super operates within a broader ecosystem that includes NVIDIA AlpaSim for closed-loop simulation, NVIDIA AlpaGym for high-throughput reinforcement learning, NVIDIA Physical AI Open Datasets for training and testing, plus open training recipes and an autolabeling pipeline. With over 500,000 downloads on Hugging Face, the platform continues to grow, inviting developers to download the model and begin building the next generation of robotaxis and autonomous vehicles.