The Intake
The Intake — Thursday, June 4, 2026
On the substrate
NVIDIA's new open-weight frontier model ships with native support for the agent platforms you're already running
If you've been watching the gap between the best open-weight models and the closed frontier, NVIDIA's Nemotron 3 Ultra is the first open-weight US model that lands inside that gap rather than below it. The weights are available today on Hugging Face, ModelScope, OpenRouter, and NIM.
The model uses a mixture-of-experts architecture. Total parameters are approximately 550 billion; active parameters per token are approximately 55 billion. NVIDIA announced the architecture at its June 1 Computex keynote. The model scores 48 on the Artificial Analysis Intelligence Index — the highest score of any US open-weight model, per The Decoder's reporting.
NVIDIA says the model delivers up to 5x faster inference and up to 30% lower cost compared with other open frontier-class models. Both figures are vendor claims; no independent methodology was published at time of writing. The model is post-trained for deployment on OpenClaw, OpenHands, LangChain Deep Agents, and Hermes Agent. If your agent stack runs on any of those platforms and you've been routing to closed APIs for frontier-class reasoning, this is the model worth benchmarking against your actual workload.
For operators
A voluntary federal pre-release review window and a new AI cybersecurity clearinghouse are now open — both relevant if your AI deployment touches federal systems or critical infrastructure
President Trump signed an executive order on June 2 establishing two voluntary structures relevant to AI operators. The first is a pre-release review window. Covered frontier AI models may be submitted for federal cybersecurity review for up to 30 days before public release. The May draft had set that window at 90 days; the signed order reduces it to 30.
The second is an AI cybersecurity clearinghouse led by the Treasury Department. NSA and CISA coordinate the effort. Participation is voluntary. Explicitly named as in-scope: rural hospitals, community banks, and local utilities. The order states that nothing in it authorizes mandatory licensing or permitting requirements.
The pre-release review framework includes a trusted partner selection mechanism — which gives the administration a role in determining who receives early model access. Policy analysts quoted by The Register characterized this as creating potential for the framework to favor politically aligned companies. That characterization is theirs, not established fact.
If your AI infrastructure already intersects with federal agency relationships, or if your deployment falls into critical infrastructure categories named in the order, the clearinghouse enrollment decision is live now. If you're a developer planning a release of a covered frontier model, the pre-release review window is the new variable in your release timeline.