← Back to feed

Nvidia introduced Alpamayo 2 Super at GTC Taipei on June 1, 2026. The model is a 32-billion-parameter reasoning-based vision language action (VLA) model for autonomous vehicle development — 3x the scale of previous Alpamayo 1 Nano and 1.5 Nano entries at 10 billion parameters each. Weights and inference code are expected on GitHub and Hugging Face this summer.

The Alpamayo family has now been downloaded close to 400,000 times since launch. The upgrade from 10B to 32B is not incremental — it changes what the model can do in production AV pipelines.

What Changed at 32B

The previous 10B Alpamayo models handled trajectory generation from video inputs. Alpamayo 2 Super extends this to a multi-task driving foundation model that covers the full autonomy stack:

• Surround-view cameras: Full 360-degree situational awareness across the vehicle’s camera array, giving the model complete context for lane changes, merges, and intersection crossing.
• Meta-Actions: High-level driving decision outputs — yield, lane change, stop — alongside the lower-level trajectories and Chain-of-Causation (CoC) reasoning traces.
• Reasoning auto-labeling with 2D grounding: The 32B foundation model can generate high-quality reasoning labels for training data, compressing annotation cycles that previously took months to days.
• Improved long-tail reasoning: Better performance on rare, complex scenarios where traditional imitation-learning AV stacks degrade.

The model is built on NVIDIA Cosmos 3 Super Reasoner, a 32B VLM backbone, and is RL post-trained.

The Teacher-Student Architecture

Alpamayo 2 Super is explicitly designed as a teacher model, not as the on-vehicle inference model. The architecture assumes distillation: the 32B foundation model provides high-quality reasoning labels and trajectory outputs, and compact derivative models are trained from those outputs to run on NVIDIA DRIVE AGX Thor — the in-vehicle compute platform.

This separates the intelligence budget from the deployment budget. A manufacturer gets frontier-scale reasoning quality without running a 32B model in a car.

Two New Open Benchmarks

Alongside the model, Nvidia released two Hugging Face challenges:

1. AlpaSim End-to-End Closed-Loop Challenge: Evaluates driving policies in closed-loop simulation, measuring time-without-at-fault incidents across reconstructed real-world scenarios.
2. Physical AI AV Reasoning Challenge: Tests model reasoning about long-tail scenarios in natural language.

Both are available now. The model weights follow this summer.

For AV developers, the calculus on Alpamayo is simple: 400,000 downloads on an open platform means the ecosystem tooling is already built. A 32B foundation model that can generate its own training labels changes the annotation economics for anyone building on it.