NVIDIA released the Alpamayo 2 Super model in 2026, with 34B parameters, comprising a 32B vision-language model backbone and a 2.3B diffusion action decoder. The model is available on the Hugging Face platform under the OpenMDW-1.1 license, supporting commercial use, fine-tuning, and redistribution of derivative models.
Model Capabilities and Benchmark Performance
Alpamayo 2 Super takes as input six surround-view camera videos, text, and ego-vehicle motion history, and outputs 64 future waypoint trajectories, causal explanation traces, meta-action labels, and visual question answering results in a single inference. It scores 79.2 on the Lingo-Judge benchmark in LingoQA, leading Qwen2.5-VL 72B by 17 points and Gemini 2.5 Pro by 15.1 points. Its closed-loop simulation AlpaSim score is 1.50±0.13, and its open-loop minimum average displacement error is 0.911 meters at 6.4 seconds.
The training data includes 115,000 hours of multi-camera driving video, over 1 billion images, and 3.7 million structured causal chain traces. These traces explicitly link observation, reasoning, and action decisions.
Practical Impact of License Changes
Early Alpamayo versions were limited to research use only. This time, the OpenMDW-1.1 license covers the entire model family, allowing developers to deploy fine-tuned models to production vehicles without additional authorization. The license terms permit retaining control over proprietary data and infrastructure.
Open-source weights enable teams to build advanced reasoning systems without training foundational capabilities from scratch, while avoiding the per-task costs of frontier closed-source models.
Substantive Divergence Between Open and Closed Camps
This release directly calls out closed-source vendors such as Anthropic, with supporters of both sides clashing over "verifiability" and "performance ceilings." Open-source advocates emphasize that full weights and causal traces can be integrated into safety verification toolchains such as NVIDIA Halos to audit the decision-making process; closed-source advocates argue that current open-source models lag behind larger closed-trained systems in generalization across extreme long-tail scenarios.
From an industry deployment perspective, Alpamayo 2 Super's open-source path lowers the entry barrier for robotaxi and long-haul trucking projects. Developers can leverage the cloud to generate synthetic trajectories and teacher outputs, then distill lightweight models suitable for on-vehicle deployment.
Deeper Driving Factors
Long-tail scenario handling is the core bottleneck in autonomous driving commercialization. Traditional perception-prediction stacks struggle to cover rare multi-agent interactions, while Alpamayo 2 Super uses reinforcement learning post-training to directly generate interpretable causal chains, attempting to transform "black-box decisions" into traceable reasoning processes. This design reflects the industry's shift from rule-driven to data-driven to verifiable reasoning-driven approaches.
NVIDIA is also opening up datasets and simulation frameworks, aiming to build a developer ecosystem around its Cosmos platform. The relaxation of commercial licensing effectively converts model capabilities into accumulable industry know-how rather than one-off API calls.
Independent Assessment
Alpamayo 2 Super's open-source commercial license already meets the conditions for practical deployment at this stage, and its benchmark data and license terms are both derived from official public materials. Future competition will center on the long-tail coverage of fine-tuned models in real vehicle fleets and the efficiency of safety verification closed loops, rather than sheer parameter scale or license format.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接