NVIDIA Nemotron 3.5 Targets Always-On Enterprise AI Agents

NVIDIA has released Nemotron 3.5, three open-weight models designed to power AI agents, pushing deeper into the shift from chatbots that answer questions to software that carries out multi-step work. Announced on October 15, 2026, the Nemotron 3.5 family is led by a 120-billion-parameter multimodal model aimed at always-on enterprise agents, alongside two smaller Bolt models for local deployment.

The headline model, Nemotron 3.5 Super VL, uses a mixture-of-experts design that activates about 12 billion parameters per token. According to NVIDIA's announcement, it handles reasoning, instruction following and tool use in a single model while understanding text, images, documents and video. The company 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.

NVIDIA's performance claims are aggressive: the company says Super VL matches the accuracy of its largest reasoning model with roughly one-fifth the parameters and delivers up to five times the throughput of comparable open models for agent work, citing Artificial Analysis leaderboard results reported by AI Parabellum.

The Bolt models bring agents to local hardware

The two smaller members of the family target deployment where the model must run on the user's own equipment. Nemotron 3.5 Bolt 74B-A9B fits in 24GB of GPU memory and is aimed at local agent workloads on RTX systems and DGX Spark hardware, while Bolt E4B, a compact 4GB model, is designed for conversational agents on devices, according to reporting by Harava News.

Both larger models are available now, while Bolt E4B is scheduled to arrive on October 23. NVIDIA says it is releasing 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.

The release fits a broader industry turn. As Adalytica's analysis notes, the economic prize of this transition is larger than a single model release: better models draw more developers and enterprises onto the platform, which in turn supports demand for the GPUs and systems that power them. NVIDIA's open models are already appearing in production enterprise agents, including Honeywell Forge Cognition, which runs operational agents on Nemotron open models for building management. For agent builders, open weights with published training recipes lower the cost of experimenting with specialized agents versus API-only alternatives.

Navan deploys Nemotron 3.5 agents in production

The models are already finding real enterprise use. On October 16, corporate travel management company Navan announced it had deployed Nemotron 3.5 Super VL in production to automate complex travel disruptions, activating 12 billion parameters per inference step across the 120-billion-parameter mixture-of-experts architecture, according to reporting by Nomad Lawyer.

The use case shows what always-on agents are for. When a business trip is disrupted by a cancellation, the problem cascades: flights need rebooking, rental cars need adjusting, hotel check-ins need changing, and every step must comply with corporate expense policies. Navan's agents execute multi-tier re-routing across flights, hotels and ground transport, parsing company spending policies in real time — moving past passive chatbots into agents that take action.

The speed of the move is notable: Navan went from initial testing to live enterprise deployment within days of receiving model access. That rapid cycle suggests open models with published recipes can compress the enterprise evaluation timeline, since teams can inspect and fine-tune the foundation rather than working through an opaque API.

There is still a burden of proof. As Spaysee's analysis of the launch points out, NVIDIA's documentation describes the training and post-training process, but the supervised fine-tuning and reinforcement learning data used for Super VL has not been publicly released, limiting outside reproduction of the company's benchmark claims. Whether these agents can identify missing evidence, obey specialized policies and explain their decisions well enough to earn trust in regulated workflows remains the decisive question.

For agent developers watching the open-model race, Nemotron 3.5 represents the clearest signal yet that the industry is converging on models designed around agentic loops — reasoning, tool use and long-context memory in one package — rather than repurposed chat models. The next test is how quickly the ecosystem builds durable, trustworthy agents on top of them.