NVIDIA Alpamayo 2 Super: The Open AV Reasoning Model That Just Beat GPT-4o by 23 Points

NVIDIA has released Alpamayo 2 Super, a fully open, commercially licensed reasoning model for autonomous vehicles — and it just ranked first on every autonomous driving benchmark NVIDIA evaluated, outscoring GPT-4o by 23.2 points and Gemini 2.5 Pro by 15.1 points on LingoQA, the leading AV reasoning benchmark.

What Alpamayo 2 Super Actually Does

Most AV perception systems answer a narrow question: what is in the scene? Alpamayo 2 Super is designed to answer a harder one: what should the vehicle do, and why? For every driving situation it processes, the model simultaneously produces five tightly coupled outputs — a planned trajectory, a chain-of-causation (CoC) reasoning trace, a meta-action label (yield, lane-change, stop), reasoning auto-labels for training data generation, and visual question-answering responses with 2D grounding that tie the model's conclusions to specific regions in camera images.

That last capability matters for safety engineering. The CoC traces integrate directly with NVIDIA Halos safety-validation workflows and support ISO/PAS 8800 alignment — the emerging standard for AI in road vehicles. Developers can audit exactly what the model observed and why it chose each action, which is a prerequisite for regulatory approval in most AV jurisdictions.

The model also ingests 360-degree camera coverage — front, sides, and rear simultaneously — giving it the full-surround situational awareness needed for complex, multi-agent scenarios like unprotected turns, lane merges at speed, and dense urban intersections. These are precisely the long-tail events where conventional AV systems most frequently fail.

Benchmark Performance: First on LingoQA Across ~40 Models

Using the Lingo-Judge evaluation metric, Alpamayo 2 Super ranks first on LingoQA — an autonomous driving reasoning benchmark — among nearly 40 models tested. NVIDIA's published margins: +17.0 points over Qwen2.5-VL 72B, +15.1 over Gemini 2.5 Pro, +23.2 over GPT-4o. The model also ranks first across all additional autonomous driving benchmarks NVIDIA evaluated, covering a broad range of AV-specific capabilities beyond pure reasoning.

At approximately 30 billion parameters — roughly 3x the scale of the earlier Alpamayo 1.5 and Alpamayo 1 models — the increased capacity helps the model generalise reasoning from sparse examples, which is the critical capability for rare, multi-agent interactions that conventional systems handle poorly.

Open Licensing Changes the Commercial Equation

Alpamayo 2 Super is available on Hugging Face under OpenMDW-1.1, the Linux Foundation's permissive licence for open AI model distributions. The licence permits fine-tuning, derivative models, and commercial redistribution without additional permissions — meaning automakers, truckmakers, Tier 1 suppliers, and AV software companies can adapt the model to proprietary fleet data, their own driving policies, and specific deployment regions, then ship it commercially.

This is strategically significant. Earlier Alpamayo releases were R&D-only; the OpenMDW licence is now being retroactively applied across the entire Alpamayo model family. AV programmes that need frontier-scale reasoning in the cloud for development and distillation — generating synthetic training data, CoC annotations, and teacher outputs — can use Alpamayo 2 Super in that role, then distil smaller, vehicle-optimised models for real-time on-board inference. NVIDIA Cosmos 3 Super Reasoner, on which Alpamayo 2 Super is built, provides the underlying multimodal foundation trained with reinforcement learning.

The Alpamayo family has already passed 500,000 downloads on Hugging Face, making it the most-adopted open reasoning model family for autonomous driving on the platform.

An Ecosystem, Not Just a Model

Alpamayo 2 Super ships alongside a full suite of complementary open tools: AlpaSim for closed-loop simulation, AlpaGym for high-throughput RL training, NVIDIA Physical AI Open Datasets, open training recipes, and an autolabeling pipeline. Together these compress annotation cycles from months to days — a model running against a raw fleet clip can generate structured CoC labels and 2D-grounded question-answering data automatically, feeding the next training run.

The Alpamayo ecosystem represents NVIDIA's clearest statement yet that it intends to own the AV AI software layer the same way it owns the GPU compute underneath it. For the hardware investment enabling this model's deployment at scale, see our coverage of Meta and BlackRock's $14B data center deal and Etched's purpose-built inference silicon. For the broader open-model competitive landscape Alpamayo is entering, see our analysis of DeepSeek V4-Flash and OpenAI's full-stack strategy. The agentic reasoning capabilities that make Alpamayo possible also underpin Oracle's enterprise AI platform.

References

• NVIDIA Blog — Alpamayo 2 Super Now Available (Aug 2026)
• Hugging Face — nvidia/Alpamayo2-Super (OpenMDW-1.1)
• NVIDIA Cosmos — Foundation Model Platform
• NVIDIA Halos — AV Safety Validation
• NVIDIA DRIVE — Open Models & Datasets
• Wired — NVIDIA Wants to Own Every Chip Inside AI Data Centers
• TechCrunch — Uber's New Robotaxi from Lucid and Nuro (2026)
• TechCrunch — Uber and Nuro Testing Premium Robotaxi in SF
• NVIDIA Newsroom — AI Factories & Data Centers
• YouTube — NVIDIA Alpamayo 2 Super Official Demo