Nvidia Debuts Nemotron 3 Ultra, Fast Open-Weight AI But China Still Leads

Nvidia unveiled Nemotron 3 Ultra at Computex 2026, a 550‑billion‑parameter open-weight model that uses mixture‑of‑experts to run with 55 billion active parameters and achieve over 300 tokens per second—three to six times faster than Chinese rivals. Independent testing puts its Intelligence Index at 48, behind Moonshot AI’s Kimi K2.6 at 54, underscoring that China currently leads the open-weight frontier. Nvidia is publishing the weights and training recipes, shipping Ultra on June 4, and has formed the Nemotron Coalition with eight labs to co-develop open frontier models on DGX Cloud. Ultra also features a 1‑million‑token context window and multi-token prediction to speed generation, reflecting Nvidia’s push to close the gap with Chinese models while expanding access via API.