NVIDIA’s 13 New Models

Sam Witteveen covers NVIDIA’s CES 2026 keynote, where Jensen Huang announced 13 open-source models across autonomous vehicles, robotics, healthcare, and enterprise AI — the first CES in five years where NVIDIA made no consumer GPU announcement, signaling a strategic pivot toward software, inference infrastructure, and developer tooling.

The hardware centerpiece was the Vera Rubin AI supercomputer platform, offering five times the inference throughput of Blackwell for data center workloads, with confirmed commitments from AWS, Google, Microsoft, and Oracle. NVIDIA also announced the platform’s improved efficiency for training Mixture of Experts models. Among the 13 models, Witteveen focuses most closely on the Neatron family: a multimodal RAG embedding model (text and vision), a multimodal reranker, and the Neatron Speech Streaming ASR model. The speech model targets low-latency real-time applications — voice agents, live captioning, in-car assistants, and edge AI — using a novel cache-based architecture that avoids redundant audio frame computation common in traditional streaming approaches.

Other announced models include AlpaMayo, a 10B open reasoning model for self-driving cars already slated for a Mercedes vehicle launch; Cosmos 2 world models for generating synthetic robotics training data; Isaac for humanoid robot control; and four Clara models for drug discovery. All models are available on Hugging Face, and Witteveen positions the Neatron Speech model as a direct successor to NVIDIA’s Parakeet ASR series.

📺 Source: Sam Witteveen · Published January 06, 2026
🏷️ Format: News Analysis