NVIDIA Releases Alpamayo 2 Super, a 34B Vision-Language-Action Model for Autonomous Driving
NVIDIA has released Alpamayo 2 Super, a 34B-parameter foundation model for autonomous vehicle development that combines a 32B vision-language backbone with a 2.3B-parameter diffusion action decoder. The model handles trajectory prediction, visual question answering, 2D grounding, and auto-labeling, and posts a Lingo-Judge score of 79.2 on LingoQA reasoning evaluation.
NVIDIA has released Alpamayo 2 Super, a 34-billion-parameter foundation model built for autonomous vehicle (AV) development, on Hugging Face. The model is part of NVIDIA's broader Alpamayo Open Platform and targets a single unified checkpoint for perception, planning, and decision-making tasks in AV software.
Architecture and Specs
Alpamayo 2 Super is a Vision-Language-Action (VLA) model built on two components: a 32-billion-parameter VLM backbone derived from Cosmos 3 Super Reasoner, and a 2.3-billion-parameter diffusion-based action decoder. The model ingests multi-camera image/video streams, text, and egomotion history (3D translation plus 9D rotation matrices across multiple timesteps), and outputs either text (Chain-of-Causation reasoning traces, meta-actions, visual question answers, grounding coordinates) or trajectories.
The trajectory API returns 64 waypoints spanning 0.1 to 6.4 seconds at 0.1-second intervals, each with ego-frame XYZ positions and 3x3 rotation matrices. Validated public notebook profiles use six cameras and four historical frames per camera.
Training Data
According to NVIDIA, the model was trained on roughly 115,000 hours of multi-camera driving video with egomotion and trajectory annotations, plus approximately 3.7 million Chain-of-Causation reasoning traces linking driving decisions to causal factors. Training data included more than 1 billion images and between 10,000 and 1 million hours of video, collected via vehicle sensors (RGB cameras, IMUs, GPS) and labeled through a hybrid automated/manual process.
Benchmark Results
NVIDIA reports the following evaluation scores:
- LingoQA reasoning: Lingo-Judge Score of 79.2
- AlpaSim closed-loop evaluation (913 scenarios from PhysicalAI-AV-NuRec Dataset): AlpaSim Score of 1.50 ± 0.13
- Open-loop evaluation (1,434 challenging samples from PhysicalAI-AV Dataset): minADE_6 at 6.4s of 0.911 meters
These figures come directly from NVIDIA's model card; no independent third-party verification is available at this time.
Licensing and Deployment
Model weights are released under the OpenMDW-1.1 license, while source code is licensed under Apache 2.0. The model has been tested on NVIDIA H100 80GB HBM3 GPUs running Linux, requiring PyTorch 2.8+, Hugging Face Transformers 4.57.1+, and DeepSpeed 0.17.4+. A single-GPU inference configuration (seven cameras, four frames each, batch size 1, BF16, 10 diffusion steps) measured a peak device memory footprint of 72,115 MiB. Pricing is not applicable, as the model is distributed as open weights rather than a hosted API product.
What This Means
Alpamayo 2 Super consolidates multiple AV development tasks — trajectory prediction, visual question answering, 2D grounding, and auto-labeling — into a single foundation model rather than requiring separate specialized systems. This mirrors a broader industry shift toward unified vision-language-action models for robotics and autonomous systems, following similar architectural bets from other AV and robotics labs.
The reported benchmark numbers (LingoQA, AlpaSim, minADE_6) are useful as directional indicators but come solely from NVIDIA's own evaluation; independent replication will determine how the model performs against competing AV foundation models in real-world deployment. The open-weight release under OpenMDW-1.1, combined with Apache 2.0 source code, positions this as an accessible baseline for AV researchers rather than a closed commercial product, though the H100-only validation and heavy memory requirements (nearly 70 GiB per GPU) mean deployment remains limited to well-resourced teams for now.
Related Articles
Nvidia and Palantir Deploy AI to Manage Supply Chains, Starting With Nvidia's Own 1.3-Million-Part Racks
Nvidia and Palantir are integrating Nvidia's open Nemotron models and cuOpt optimization engine into Palantir's Foundry platform to manage complex supply chains. The first deployment is Nvidia's own operation, where a single Vera Rubin server rack contains 1.3 million components.
DeepSeek Releases V4.1-Flash: 552B MoE Model Cuts KV Cache to 890 Bytes Per Token
DeepSeek has released V4.1-Flash, a 552B-parameter multimodal Mixture-of-Experts model supporting 1M-token context and activating only 8B parameters during prefill. The model uses a new Causal Encoder-Decoder architecture and Compressed Sparse Attention 2 to cut global KV cache to 890 bytes per token, roughly a quarter of its predecessor.
Alibaba Open-Sources Qwen3.8-2.4T-A95B, Its First Qwen-Max-Class Model With Public Weights
Alibaba's Qwen team released Qwen3.8-2.4T-A95B on August 12, 2026, the open-weight version of Qwen3.8-Max and the first Qwen-Max-class model made publicly available. The 2.4 trillion-parameter mixture-of-experts model activates only 95 billion parameters per token and supports context windows up to 1 million tokens.
AllSpark's Iris-mini and Iris-pro Top Open-Weight Search Agent Benchmarks
Chinese lab AllSpark has released Iris-mini and Iris-pro, two open-weight search agents built on Qwen3 models that claim the top spot among open-weight systems in their size classes on four research benchmarks. The release includes model weights, an agent harness, and evaluation code, with training pipelines to follow.
Comments
Loading...