NVIDIA Alpamayo 2 Super, the Frontier Open Model for Robotaxis and Autonomous Vehicles, Now Available for Commercial Use
NVIDIA Jessica Soares
Nvidia just open-sourced Alpamayo 2 Super, a big new AI brain for self-driving cars, free for commercial use. It beats Gemini, GPT-4o and Qwen on driving reasoning tests, which is a genuinely big deal for robotaxi makers.
Nvidia has a habit of open-sourcing the unglamorous plumbing that everyone else quietly needs, and Alpamayo 2 Super fits that pattern. It's the newest entry in the Alpamayo family, a line of AI models built specifically to help autonomous vehicles handle the weird stuff — the unprotected left turn with a jaywalker, the merge where three cars all hesitate at once. Those long-tail scenarios are where most self-driving stacks quietly fall apart, and Nvidia is betting that a bigger, better-reasoning model can close that gap.
The headline number is 30 billion parameters, three times the size of the earlier Alpamayo 1 and 1.5 models, and Nvidia says the extra scale pays off directly on benchmarks. On LingoQA, a widely used driving-reasoning test, Alpamayo 2 Super topped nearly 40 competing models, beating Gemini 2.5 Pro by 15.1 points, GPT-4o by 23.2 points and Qwen2.5-VL 72B by 17.0 points using Nvidia's own Lingo-Judge metric. It also reasons across a full 360-degree camera rig rather than just a forward-facing view, which matters a lot when the danger is coming from the side mirror, not the windshield.
What makes this more than a leaderboard flex is the licensing. Nvidia is releasing the whole model under OpenMDW-1.1, a permissive license from the Linux Foundation that allows fine-tuning, derivative models and outright commercial redistribution — no extra permission slip required. Automakers, robotaxi operators and suppliers can now take Alpamayo, retrain it on their own proprietary fleet data, and ship it in production vehicles. That's a meaningfully different posture than most frontier AI releases, which tend to stay locked behind research-only terms until a company decides it's ready to monetize.
The model doesn't just spit out a steering path either. For each driving decision it generates a planned trajectory, a plain-language explanation of why it chose that action, a labeled intent like
My take
Nvidia open-sourcing a frontier driving model under a genuinely commercial license is the kind of move that should embarrass every AV startup still hoarding mediocre in-house models behind NDAs. If a 30B open model can out-reason GPT-4o on driving tasks, the moat for most robotaxi companies was never the AI brain, it was the fleet data and the willingness to actually ship — and now Nvidia just made the excuse of building your own reasoning stack from scratch look pretty silly.
Read more about this at: NVIDIA
