Nvidia releases Nemotron 3.5 Super VL, a 120-billion-parameter open-weight model for long-running agents

Nvidia yesterday released Nemotron 3.5 Super VL, an open-weight multimodal model built to power AI agents that operate over extended periods. The model fuses text, image, document and video understanding in a single architecture, eliminating the need to shuttle context between a separate language model and a separate vision model — a change intended to cut latency and simplify agent system design.

MoE architecture: 120 billion total, 12 billion active

Under the hood sits a mixture-of-experts design: 120 billion parameters in total, but only 12 billion active per forward pass. The practical upshot is substantially lower inference cost than a dense model of the same nominal size while retaining high representational capacity. Nvidia did not publish benchmarks against rivals such as Llama 3.1 405B or its own prior Nemotron 3 Ultra, leaving the efficiency-quality trade-off open to independent verification.

Native multimodal capabilities: documents, charts, video

The model was designed from the ground up to ingest full documents, interpret charts and tables, and process video sequences — capabilities that normally require a pipeline of several specialist models. According to Nvidia, this unification lets agents read financial reports, analyse UI screenshots and browse the web in a single pass without losing cross-modal context. The company showed a chart claiming accuracy gains after post-training for specific tasks, but did not disclose sample size, evaluation sets or measurement protocol.

Post-training for verticals: cyber, finance, science, travel

External teams are already running post-training on top of Nemotron 3.5 Super VL for five domains: document intelligence, cybersecurity, finance, scientific experimentation and travel planning. The pattern — strong general base, then narrow specialisation — is the industry standard, though Nvidia has not said whether the recipes and data used for that post-training will also be released under the same open licence or remain proprietary to the customers.

"Open" in Nvidia's sense: weights, data, recipes

As with earlier Nemotron family releases, the model ships with weights, training data and training recipes — what Nvidia calls "open and customisable." In practice the usage licence may restrict commercial use or derivative distribution; the full licence text was not attached to the announcement. The distinction between "open weights" and "open source" in the full OSI sense remains relevant, and developers should review the terms before embedding the model in a product.

Availability and technical details

The model is available starting today for download and local or cloud deployment; full technical details — including the post-training approach, multi-token prediction and inference optimisations — are documented in Nvidia's technical post at nvda.ws/3Wqw7i6. In the absence of independent benchmarks and real-world stress tests, the community will need to run its own evaluations before deciding whether Nemotron 3.5 Super VL replaces the existing model chain in production environments.