Nvidia opens its largest self-driving model to commercial fine-tuning

Nvidia released Alpamayo 2 Super today under a Linux Foundation license that lets robotaxi and autonomous-vehicle developers fine-tune the model, build derivatives, and ship commercial products without negotiating a separate deal. That is a meaningful shift for a model that had been positioned for research and development use only. It is also the most concrete signal yet of how Nvidia plans to keep its grip on self-driving infrastructure: give away the weights, lock the stack.

Alpamayo 2 Super is a self-driving AI model, the kind of software that turns camera and sensor feeds into driving decisions. It is built on Nvidia's Cosmos 3 Super Reasoner and post-trained with reinforcement learning, a trial-and-error training method that lets the model improve from driving outcomes rather than just labelled examples. It is roughly three times the size of Nvidia's 10-billion-parameter Alpamayo 1.5 and Alpamayo 1, and Nvidia is positioning it as the largest option in a three-tier family aimed at the full deployment pipeline: Alpamayo 2 Super for cloud-based reasoning and development, Alpamayo 1.5 and 1 for cheaper development and distillation (training a smaller, faster model to mimic the larger one's outputs), and distilled variants optimised for real-time in-vehicle use.

"Open weights" here means the trained model parameters are downloadable, so a developer can run and adapt the model on their own infrastructure. It does not mean the training data, the training code, or the full development pipeline is open in the way open-source software is. The model runs and the weights are yours; the recipe is not.

The new license is OpenMDW-1.1, a Linux Foundation framework. It allows fine-tuning, derivative models, and commercial redistribution, the permissions that matter most to a robotaxi operator or a Tier 1 supplier that wants to ship a derivative model in production. Nvidia says the same open-licensing approach is being extended across the wider Alpamayo family. The Hugging Face model card for Alpamayo 2 Super and Nvidia's developer blog post lay out the terms.

That is the part of the story a standard "Nvidia open-sources self-driving model" headline misses. The license is a Linux Foundation license, which is a real third-party governance structure rather than a Nvidia-proprietary one. But every Alpamayo model runs on Cosmos, fine-tunes against Nvidia's toolchain, and ships into vehicles on Nvidia's in-vehicle compute. The lock-in is not in the legal terms; it is in the deployment pipeline around them.

The benchmark case Nvidia is making for Alpamayo 2 Super is, by the company's own framing, an Nvidia-run comparison. In Nvidia's testing on its own Lingo-Judge metric, the company says the model beat Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points. Nvidia also says it ranked first on LingoQA, a self-driving reasoning benchmark, among nearly 40 models the company evaluated. Lingo-Judge is an internal Nvidia metric. These are Nvidia's numbers, and the comparison set is Nvidia's choice. MarkTechPost, TheNextWeb, and IoT Tech News have all repeated the same figures without independent verification.

The lock-in matters more the longer a vehicle program runs. An AV developer that fine-tunes an Alpamayo variant for a commercial launch is choosing a pipeline they will live inside for the lifetime of that program, which is exactly the kind of long-tail dependency Nvidia is buying by giving away the weights today.

The next thing to watch is whether any major AV developer publicly accepts the license and ships a derivative model, or whether teams that have built around closed-frontier self-driving stacks treat the Linux Foundation framing as sufficient. The terms on the page are permissive. The stack around them is Nvidia's.