NVIDIA releases Nemotron 3.5 Super VL and new Bolt models to expand its open, agent-focused LLM lineup for on-premises and local deployment

Decision brief

What changed

NVIDIA announced new open Nemotron models for agentic deployments: Nemotron 3.5 Super VL, plus Nemotron 3.5 Bolt 74B-A9B and Bolt E4B variants. The company said the models will be available as downloadable weights under an open commercial license so organizations can run and post-train them on their own infrastructure, including local and edge devices with 24GB or as little as 4GB of memory.

Why it matters

This expands the practical options for enterprises that want agent-capable and multimodal models without depending only on hosted APIs, especially where data-control, latency, or offline deployment requirements matter. For decision-makers, the main implication is a wider design space between data-center and endpoint deployment, with NVIDIA explicitly targeting customizable on-premises and local execution rather than only cloud consumption. If the stated memory footprints and enterprise licensing terms hold in production, teams could evaluate lower-infrastructure-footprint pilots for internal agents and edge use cases.

Affected roles

CEO COO CTO CISO

Evidence

The event is primarily supported by NVIDIA’s own announcement, which describes the model releases, open weights, deployment positioning, and claimed efficiency. Two independent trade outlets, The New Stack and SiliconANGLE, consistently corroborate the core facts on model types, local-memory targets, commercial/open licensing, and on-premises or behind-firewall deployment, though performance claims remain company-sourced.

What remains uncertain

Open questions include real-world performance versus alternatives, total hardware and operational requirements beyond the cited memory footprints, and how much post-training/customization effort enterprises will need to achieve production results. It is also not fully verified in the coverage how broadly the licensing, support tooling, and availability timelines will meet enterprise needs; this brief assumes the downloadable open-weight access and commercial-use terms are implemented as described.

Monitor next

Watch for independent benchmarks, general availability of the weights on Hugging Face, and early enterprise reports on whether the 24GB and 4GB Bolt variants perform acceptably in real agent workflows.

Analytical support, not advice — assumptions and open questions stated above.

Source coverage