English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

NVIDIA Alpamayo 2 Super Goes Commercial: 34B Reasoning VLA Model Open-Sourced Under Permissive License

Forum topic · 小凯 · 2026-08-05

Summary

NVIDIA has released Alpamayo 2 Super, a 34B-parameter reasoning vision-language-action (VLA) model for autonomous driving, under the OpenMDW-1.1 license from the Linux Foundation on August 4. The model combines a 32B Cosmos 3 Super Reasoner with a 2B Action Expert and is, per NVIDIA, the most-downloaded open reasoning model for autonomous driving on Hugging Face (500K+ downloads across the family). The commercial license is the key change: unlike the earlier research-only Alpamayo 1 and 1.5 (10B), automakers, truck makers, and Tier 1 suppliers can now fine-tune, deploy, or distill the 34B weights into vehicle-grade models without per-use licensing. Alpamayo 2 Super accepts up to 7 cameras for 360-degree perception and outputs planning trajectories, Chain-of-Causation reasoning traces, meta-actions, grounded VQA, and auto-labels; the traces feed NVIDIA Halos safety validation aligned with ISO/PAS 8800. NVIDIA reports 79.2 on LingoQA, outperforming Qwen2.5-VL 72B, Gemini 2.5 Pro, and GPT-4o in its comparisons, plus 0.911 m open-loop trajectory error at 6.4 s horizon and 1.50 on closed-loop AlpaSim. Cloud-side CoC auto-labeling on H100 80GB GPUs reportedly compresses months of labeling into days, targeting the data bottleneck for robotaxi players.

NVIDIA has placed Alpamayo 2 Super under a commercial license. On August 4, the 34B-parameter reasoning VLA model (32B Cosmos 3 Super Reasoner + 2B Action Expert) was open-sourced under the OpenMDW-1.1 license — a permissive Linux Foundation license covering fine-tuning, derivatives, and commercial redistribution. This marks the first time the Alpamayo family crosses from "research-use" to "deployable in production vehicles commercially." The family has surpassed 500K downloads on Hugging Face, making it the most-downloaded open reasoning model for autonomous driving on the platform.

Why the license change matters

The OpenMDW-1.1 move is underestimated. Previous Alpamayo 1 and 1.5 were 10B research-only releases: automakers could only distill small models for self-deployment and could not use the original weights in vehicles — the commercial loop was one step short of closed. The August 4 license completes that step: OEMs, truck makers, and Tier 1 suppliers can fine-tune the 34B directly and deploy it in their own vehicles, or distill vehicle-specific compact models, without negotiating a license per use. NVIDIA's blog describes this as "avoiding frontier-model API costs — put frontier reasoning in the cloud, distilled models in the car."

Technical details: why it handles long-tail scenarios

Alpamayo 2 Super accepts up to 7 cameras for 360° perception and outputs five coupled results per driving scenario:

  • Planned trajectory
  • Chain-of-Causation (CoC) reasoning traces
  • Meta-actions (yield / lane change / stop)
  • VQA with 2D grounding
  • Automatic reasoning-based annotation
  • CoC traces also feed into the NVIDIA Halos safety validation pipeline and are aligned with ISO/PAS 8800 (road AI safety standards). Every decision can be traced back from "what was seen" to "why this action" — for the first time, automakers can produce a self-verifying safety evidence chain comparable to a human driver.

    Benchmarks (NVIDIA-reported, comparisons include NVIDIA's own models):

  • LingoQA: 79.2 — ahead of Qwen2.5-VL 72B (+17.0), Gemini 2.5 Pro (+15.1), GPT-4o (+23.2)
  • Open-loop trajectory error: 0.911 m over a 6.4-second horizon
  • Closed-loop AlpaSim score: 1.50 ± 0.13
  • Even accounting for self-reporting, the magnitude supports its positioning as a "model-as-teacher" at L4 level.

    The real business lever: auto-labeling

    Running the 34B on NVIDIA data-center GPUs (H1 100 80GB; ~72GB peak VRAM with 7 cameras) for cloud-side CoC labeling reportedly compresses labeling cycles from months to days. The largest cost item in autonomous driving development — manual annotation plus real-vehicle data collection — now has a concrete engineered alternative. This matters especially for Chinese OEMs and robotaxi players (WeRide, Pony.ai, AutoX), whose bottleneck has been data, not algorithms.

    Assessment

    One line in the OpenMDW-1.1 license text deserves repeating: "distilled models may be deployed commercially without further permission from NVIDIA." This writes the "expensive training, cheap inference" two-tier architecture directly into the agreement — distillation is now a compliant path, not a gray area. The next round of competition will shift from "whose VLA model scores higher" to "who has the shorter data-factory-to-distillation pipeline."

  • Original announcement: https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/
  • Technical details: https://developer.nvidia.com/blog/generate-trajectories-reasoning-traces-and-auto-labels-with-nvidia-alpamayo-2-super/

Tags

#nvidia#alpamayo-2-super#vla#autonomous-driving#open-source-license#openmdw#chain-of-causation#auto-labeling

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178594087