NVIDIA Expands Nemotron 3 With Speech and RAG Models

NVIDIA broadened its Nemotron 3 model family on June 30, adding open-weight models for speech and multimodal retrieval, per the NVIDIA Blog.

For builders who can't send data to a hosted API, this is the first concrete on-prem path. It bundles open weights, an inference server with day-0 support, and a sovereign hosting partner into one release.

Palantir launched an engine that runs Nemotron inside air-gapped US government environments. Agencies barred from cloud AI can deploy the same models NVIDIA ships publicly, without the data ever leaving their own network.

Nemotron 3 Ultra, the largest model in the family, got day-0 support in vLLM, the open-source inference server. That matters because the biggest model in a release is often the last one anyone can actually run. Here it's usable from day one.

The update also shipped open datasets and libraries aimed at agentic AI, plus new safety models. It's a release built for teams assembling their own stack rather than calling someone else's API.

A deployable stack, not a spec sheet

Open weights, vLLM, and sovereign hosting used to be three separate bets a team had to place and hope they'd line up. NVIDIA and Palantir wired them together into one deployable path. Regulated teams that couldn't touch a hosted API have somewhere to go. Palantir's air-gapped engine is the part that will pull sovereign AI work away from cloud vendors first, not last.

NVIDIA dated the release June 30, covered here in the July 1 briefing.