NVIDIA teases Nemotron 3 Ultra, a 550B-parameter MoE model for AI agents

NVIDIA's official AI account teased that "Nemotron 3 Ultra is coming this week," with reports pegging availability for June 4. The model is a 550-billion-parameter mixture-of-experts (MoE) system explicitly built for AI agents — a sparse architecture that activates only a subset of parameters per token to balance capability against inference cost, which matters for the long-running agentic workloads NVIDIA is targeting.

Nemotron 3 Ultra rounds out a busy GTC Taipei / Computex week for NVIDIA that also included Cosmos 3 (physical AI), JetPack 7.2 (edge), and the RTX Spark superchip (consumer/prosumer). Where Cosmos handles embodied world models, Nemotron targets text-and-reasoning agentic workloads, giving NVIDIA an open frontier-scale model to pair with its hardware and tooling.

Releasing a 550B open MoE positions NVIDIA against open-weights efforts from Alibaba's Qwen, DeepSeek, Meta's Llama line, and MiniMax — while reinforcing NVIDIA's posture as a major open-source AI contributor, a reputation Hugging Face's Clément Delangue publicly endorsed this week. Crucially, an NVIDIA-built flagship model also showcases what its own silicon can do, a self-reinforcing demand loop.

Because the model isn't out yet, key questions remain: active-parameter count, context length, licensing terms, and how it benchmarks against Qwen3.7-Max and GPT-5.5 on agentic and coding tasks. Watch the June 4 release and accompanying benchmarks to judge whether Nemotron 3 Ultra is competitive at the frontier or primarily a hardware demonstration vehicle.