NVIDIA Releases Alpamayo 2 Super: A 34B Open Vision-Language-Action Model for Robotaxis and Autonomous Driving Under OpenMDW-1.1
MarkTechPost Asif Razzaq ● Covered by 2 sources
NVIDIA open-sourced Alpamayo 2 Super, a 34B self-driving AI you can actually use commercially, no strings attached. It beats Gemini and GPT-4o on driving reasoning tests, and explains its decisions instead of just guessing.
NVIDIA just handed the autonomous vehicle world a fairly serious gift: a 34-billion-parameter vision-language-action model called Alpamayo 2 Super, released under a license that actually lets you ship it. Not a research toy you admire and then can't touch commercially. OpenMDW-1.1, backed by the Linux Foundation, covers fine-tuning, derivatives, and redistribution from the moment you download it, and NVIDIA retroactively applied the same terms to earlier Alpamayo releases that used to require special permission.
The architecture pairs a 32B Cosmos 3 Super Reasoner backbone, tuned with reinforcement learning, with a smaller 2.3B diffusion decoder that turns reasoning into an actual trajectory. Feed it six camera feeds and four frames of history per camera, plus egomotion data, and it spits out 64 waypoints stretching from 0.1 to 6.4 seconds ahead. That's the easy part, honestly. What makes this one interesting is that the same forward pass also produces a plain-language explanation of why it chose that path, a meta-action label like 'yield' or 'lane change,' and grounded visual question answering. One model, five things out.
The training diet is enormous by any standard: roughly 115,000 hours of multi-camera driving footage, over a billion images, and 3.7 million so-called Chain-of-Causation traces that link a decision to its cause. That last part is the piece NVIDIA is clearly betting on. Regulators and safety engineers don't just want a car that swerves correctly, they want to know why it swerved, and CoC traces are built to slot directly into NVIDIA's Halos safety validation pipeline aligned with ISO/PAS 8800.
On benchmarks, the numbers are hard to wave away. Alpamayo 2 Super posted a Lingo-Judge score of 79.2 on LingoQA, putting it first out of nearly 40 models and beating Qwen2.5-VL 72B by 17 points, Gemini 2.5 Pro by 15, and GPT-4o by more than 23. In closed-loop simulation across 910 scenarios it scored 1.50 ± 0.13, and its open-loop minADE at 6.4 seconds came in at 0.911 meters on nearly a thousand hard cases. None of that means it's ready to steer your commute tomorrow — this thing peaks at 72GB of memory on a single H100, so it's a cloud-scale reasoning engine that needs distilling down before it lands in a car's compute stack. But as an autolabeler for fleets that are drowning in unannotated footage, NVIDIA claims it turns months of labeling work into days, which is arguably the more immediate business case than robotaxis themselves.
My take
Open-sourcing a model this capable under a genuinely permissive license is the smart move, since the real bottleneck in self-driving was never algorithms, it was labeled data and liability-proof explanations, and NVIDIA just gave away tools for both. The benchmark flex against GPT-4o and Gemini is nice marketing, but the boring detail about compressing annotation cycles from months to days is the part that will actually move the industry forward.
Read more about this at: MarkTechPost
