NVIDIA Alpamayo 2 Super AI generating trajectories and reasoning traces, showcasing advanced computing power.

Editorial illustration for NVIDIA's Alpamayo 2 Super Generates Trajectories and Reasoning Traces

NVIDIA's Alpamayo 2 Super Generates Trajectories and...

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

August 5, 2026 • 3 min read

Autonomous vehicle teams typically juggle a separate model for trajectory prediction, another for intent forecasting, a third for scene understanding, and yet another for labeling raw sensor data. That fragmentation makes it hard to trace why a model made a decision or to reuse the same outputs across different stages of development. NVIDIA is trying to close that gap with Alpamayo 2 Super, a 34-billion-parameter reasoning vision-language-action model built for AV work.

It pairs a 32-billion-parameter Cosmos 3 Super Reasoner with a 2-billion-parameter diffusion-based Action Expert, then runs reinforcement learning post-training on top. The reasoner reads multi-camera video, language prompts, and past motion, while the Action Expert turns that internal representation into a projected ego-vehicle path.

Coverage now spans 360 degrees across as many as seven cameras, and the model outputs more than a single trajectory: reasoning traces, meta-actions, scene-grounded answers, and auto-generated labels all come from the same foundation. That means one model can serve as a policy teacher, an evaluation critic, or a data engine, rather than requiring a different network for each job. NVIDIA outlines four hands-on workflows built around this setup, starting with trajectory and reasoning-trace generation.

Alpamayo 2 Super brings surround-view perception, reasoning, planning, scene understanding, and data auto-labeling into one open model workflow. These capabilities give developers a practical foundation for building teacher models, curating long-tail data, inspecting policy decisions, and evaluating AV systems beyond a single open-loop trajectory metric.

Generate Trajectories, Reasoning Traces, and Auto-Labels with NVIDIA Alpamayo 2 Super - NVIDIA Developer Blog

Why this matters

For anyone building AV stacks, the pitch here isn't a marginal accuracy bump, it's consolidation. NVIDIA is betting that folding trajectory generation, intent prediction, scene reasoning, and auto-labeling into one 34-billion-parameter model solves a real workflow problem: teams currently juggle separate models that don't share representations, which makes debugging and comparing outputs a mess. If Alpamayo 2 Super actually delivers a shared backbone across those four tasks, that's a legitimate engineering win, fewer handoffs, easier to trace why a model made a call.

But "open" and "34B params" invite the usual questions we should keep asking: what's the actual compute cost to run this in production, how good are the auto-labels versus human-verified ground truth, and does reasoning-trace output translate into safer decisions or just more plausible-sounding justifications after the fact. The code snippet suggests a fairly accessible API, which matters for researchers who want to poke at it directly rather than take NVIDIA's benchmarks on faith. Worth watching whether independent labs replicate the claimed gains before this becomes the default AV foundation model.