NVIDIA Launches Cosmos 3, Alpamayo 2, Vera Rubin, Acquires Kumo AI, and Pushes RTX Spark for Local AI Agents
Key Questions
What is NVIDIA Cosmos 3?
Cosmos 3 is an open world foundation model for physical AI using mixture-of-transformers architecture. It offers Super/Nano/Edge tiers and a Cosmos Coalition with industry partners.
What is RTX Spark and its capabilities?
RTX Spark enables local AI agent PCs with 1000 TOPS and 128GB unified memory, supporting 120B models on laptops. It delivers 2x llama.cpp gains and partnerships with Microsoft, Dell, and HP.
What is Alpamayo 2?
Alpamayo 2 is a 32B open VLA model for robotaxis with full-stack reasoning and closed-loop RL via AlpaGym. It builds on prior NVIDIA robotics releases.
What did NVIDIA acquire and why?
NVIDIA acquired Kumo AI for over $400M to expand enterprise model-making capabilities. It also partners with Palantir to integrate Nemotron.
What is Vera Rubin and its performance?
Vera Rubin is NVIDIA's data center platform now in full production, delivering 10x agent throughput over Grace Blackwell. It uses CPO-based Spectrum-X Ethernet.
What open models did NVIDIA release?
Nemotron 3 Ultra is a 550B MoE open-weight model optimized for long-running agents and available on Hugging Face. It adopts multi-teacher on-policy distillation for post-training.
How does RTX Spark support local agents?
RTX Spark powers local agent deployment with OpenShell and multi-GPU tensor parallelism. HP showcased compatible AI-ready PCs at Computex 2026.
What research support does NVIDIA provide for physical AI?
NVIDIA supplies agentic skills for autonomous vehicles, robotics, and vision AI to automate simulation and policy training. A high-concurrency serving bug was noted for Nemotron API.
NVIDIA launches Cosmos 3 open world foundation model for physical AI (mixture-of-transformers, two-tower reasoner VLM + diffusion generator, Super/Nano/Edge tiers, Cosmos Coalition with industry partners). Simultaneously pushes RTX Spark CPU for AI agent PCs with Microsoft, Dell, HP partnerships, enabling local agent deployment with OpenShell, 2x llama.cpp gains on Qwen 3.6, and multi-GPU tensor parallelism. RTX Spark offers 1000 TOPS, 128GB unified memory, enabling 120B models on thin-and-light laptops. HP announces RTX Spark PCs at Computex 2026, expanding local AI agent hardware ecosystem. New: Alpamayo 2 Super open VLA model for robotaxis (32B, full-stack reasoning, meta-actions, closed-loop RL via AlpaGym). New: 4D-RGPT open-source model for native 4D spatiotemporal understanding built on alpamayo-R1. Vera Rubin data center platform enters full production (10x agent throughput over Grace Blackwell, CPO-based Spectrum-X Ethernet). Also partnering with Palantir to integrate Nemotron into enterprise agent platform. Acquires Kumo AI for $400M+ to expand enterprise model-making capabilities. Cosmos 3 paper released detailing omnimodal world model architecture. NVIDIA also provides agentic skills for physical AI research (AV, vision, robotics) to automate simulation, data generation, and policy training. Nemotron 3 Ultra (550B MoE, open-weight) now available on Hugging Face, optimized for long-running agents. Industry standard post-training: NVIDIA adopts multi-teacher on-policy distillation (MODP). New: A high-concurrency serving bug for Nemotron 3 Ultra reported via API, with local runs showing no issues – affects benchmark trustworthiness and deployment decisions.