NVIDIA Launches Nemotron 3.5: Open Models Built for Always-On Enterprise Agents
NVIDIA released three open-weight Nemotron 3.5 models for AI agents, led by a 120B multimodal Super VL built for multi-day enterprise workflows.
What happened
On October 15, 2026, NVIDIA released Nemotron 3.5 — three open-weight models designed specifically to power AI agents, pushing deeper into the industry's shift from chatbots that answer questions to software that carries out multi-step work.
The family is led by Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model using a mixture-of-experts design that activates about 12 billion parameters per token. It handles reasoning, instruction following, and tool use in a single model, understanding text, images, documents, and video. NVIDIA describes it as built for specialized agents that run for long stretches — sometimes across days or weeks — on tasks like document review, web research, coding, cybersecurity triage, and video analysis.
Two smaller Bolt models target local deployment: Bolt 74B-A9B fits in 24GB of GPU memory for local agent workloads on RTX systems and DGX Spark hardware, while Bolt E4B is a compact 4GB model for conversational agents on devices, arriving October 23.
Already in production
The models are not theoretical. On October 16, corporate travel management company Navan announced it had deployed Super VL in production to automate complex travel disruptions — flights rebooked, rental cars adjusted, hotel check-ins changed, all while parsing corporate expense policies in real time. Navan went from initial testing to live enterprise deployment within days of receiving model access.
Honeywell is also running operational agents on Nemotron open models for building management through its Forge Cognition platform.
NVIDIA is releasing the weights, datasets, and training recipes, so developers can adapt the models to their own data and workflows — a meaningful differentiator for enterprises that want agents tuned to internal rules without surrendering control of sensitive data.
What this means for your business
- The chatbot era is ending; the agent era is here. The economic prize is no longer answering questions — it is software that executes multi-step work across your systems. Ask which of your workflows could run on agents rather than people clicking through screens. - Open weights compress your evaluation cycle. Because teams can inspect and fine-tune the foundation rather than working through an opaque API, Navan-style deployments can go live in days. If your data cannot leave your environment, the Bolt models offer on-prem and on-device options with real capability. - Tune on your rules, not someone else's. Releasing training recipes lets you build agents that obey your specific policies and compliance requirements — critical for regulated industries where a generic agent's judgment is not enough. - Keep the burden of proof. NVIDIA's fine-tuning and reinforcement-learning data has not been fully released, limiting independent verification of its benchmark claims. Run your own evaluations before trusting any vendor's numbers — for agents especially, trust is earned in production, not on leaderboards.
What to watch
Whether the ecosystem builds durable, trustworthy agents on top of these models — and whether open-weight agents win in regulated industries where data control is non-negotiable.