NVIDIA CEO Jensen Huang announced the release, calling it "a powerful backbone for robotaxis, trucks, shuttles, delivery vans, tractors and the long tail of mobile robots" . The model was first shown at CES 2026 and GTC Taipei earlier that year .

Architecture: 32B Vision-Language Backbone + 2B Action Decoder

Alpamayo 2 Super totals approximately 34 billion parameters . It combines:

• A 32-billion-parameter vision-language backbone built on NVIDIA Cosmos 3 Super Reasoner .
• A 2-billion-parameter diffusion-based Action Expert (Action Decoder) .
• Reinforcement learning (RL) post-training applied to the base Cosmos 3 Super Reasoner .

The architecture is a Transformer with full-surround awareness, processing video from up to seven cameras for 360-degree coverage, and is aimed squarely at Level 4 driving .

Teacher Model: Five Coupled Outputs Per Scene

NVIDIA designed Alpamayo 2 Super as a cloud-based teacher model that produces five coupled outputs per driving scene, which can then be distilled into smaller, in-vehicle deployment models :

1. Planned trajectory — the predicted driving path.
2. Chain-of-causation reasoning trace — an interpretable reasoning chain explaining the driving decision .
3. High-level meta-action — a high-level driving action (e.g., "yield," "change lane").
4. Auto-generated reasoning labels — automatically labeled rationales for training data.
5. Visual question answers (VQA) — answers to visual queries about the driving scene .

The chain-of-causation traces integrate directly with NVIDIA Halos safety-validation workflows and support AI safety aligned with ISO/PAS 8800 requirements .

Benchmark Performance: First on LingoQA

In NVIDIA's testing, Alpamayo 2 Super achieved a LingoQA score of 79.2, ranking first on the LingoQA autonomous driving benchmark . It outperformed:

• Qwen2.5-VL 72B by 17 points
• GPT-4o by 23.2 points
• Gemini 2.5 Pro (NVIDIA's reported testing)

Strategic Positioning: Open Infrastructure for the AV Industry

NVIDIA explicitly positions Alpamayo 2 Super as open infrastructure for the entire autonomous vehicle industry — not as a competitor to robotaxi operators or AV makers . The model is available for any company to fine-tune, adapt, and commercially deploy without proprietary lock-in .

The release clears a licensing barrier that had kept the earlier Alpamayo family in research-only use. NVIDIA says the OpenMDW license now applies across the entire Alpamayo lineup, so developers can adapt the model for their own autonomous driving stacks .

Key Takeaways

• Commercial release on August 4, 2026 with weights on Hugging Face .
• OpenMDW-1.1 license permits fine-tuning, derivatives, and commercial use .
• 34B total parameters: 32B Cosmos 3 Super Reasoner + 2B diffusion Action Expert .
• Five coupled outputs per scene for cloud-based teacher model workflow .
• LingoQA 79.2 — first on the benchmark, outperforming GPT-4o by 23.2 points .
• Targets long-tail edge cases and aligns with ISO/PAS 8800 via Halos .

NVIDIA CEO Jensen Huang announced the release, calling it "a powerful backbone for robotaxis, trucks, shuttles, delivery vans, tractors and the long tail of mobile robots" . The model was first shown at CES 2026 and GTC Taipei earlier that year .

Architecture: 32B Vision-Language Backbone + 2B Action Decoder

Alpamayo 2 Super totals approximately 34 billion parameters . It combines:

• A 32-billion-parameter vision-language backbone built on NVIDIA Cosmos 3 Super Reasoner .
• A 2-billion-parameter diffusion-based Action Expert (Action Decoder) .
• Reinforcement learning (RL) post-training applied to the base Cosmos 3 Super Reasoner .

The architecture is a Transformer with full-surround awareness, processing video from up to seven cameras for 360-degree coverage, and is aimed squarely at Level 4 driving .

Teacher Model: Five Coupled Outputs Per Scene

NVIDIA designed Alpamayo 2 Super as a cloud-based teacher model that produces five coupled outputs per driving scene, which can then be distilled into smaller, in-vehicle deployment models :

1. Planned trajectory — the predicted driving path.
2. Chain-of-causation reasoning trace — an interpretable reasoning chain explaining the driving decision .
3. High-level meta-action — a high-level driving action (e.g., "yield," "change lane").
4. Auto-generated reasoning labels — automatically labeled rationales for training data.
5. Visual question answers (VQA) — answers to visual queries about the driving scene .

The chain-of-causation traces integrate directly with NVIDIA Halos safety-validation workflows and support AI safety aligned with ISO/PAS 8800 requirements .

Benchmark Performance: First on LingoQA

In NVIDIA's testing, Alpamayo 2 Super achieved a LingoQA score of 79.2, ranking first on the LingoQA autonomous driving benchmark . It outperformed:

• Qwen2.5-VL 72B by 17 points
• GPT-4o by 23.2 points
• Gemini 2.5 Pro (NVIDIA's reported testing)

Strategic Positioning: Open Infrastructure for the AV Industry

NVIDIA explicitly positions Alpamayo 2 Super as open infrastructure for the entire autonomous vehicle industry — not as a competitor to robotaxi operators or AV makers . The model is available for any company to fine-tune, adapt, and commercially deploy without proprietary lock-in .

The release clears a licensing barrier that had kept the earlier Alpamayo family in research-only use. NVIDIA says the OpenMDW license now applies across the entire Alpamayo lineup, so developers can adapt the model for their own autonomous driving stacks .

Key Takeaways

• Commercial release on August 4, 2026 with weights on Hugging Face .
• OpenMDW-1.1 license permits fine-tuning, derivatives, and commercial use .
• 34B total parameters: 32B Cosmos 3 Super Reasoner + 2B diffusion Action Expert .
• Five coupled outputs per scene for cloud-based teacher model workflow .
• LingoQA 79.2 — first on the benchmark, outperforming GPT-4o by 23.2 points .
• Targets long-tail edge cases and aligns with ISO/PAS 8800 via Halos .