NVIDIA Releases Nemotron 3.5 Super VL for Multimodal AI Agents

NVIDIA released Nemotron 3.5 Super VL on Thursday, Oct. 15, targeting organizations building AI agents that process different media and handle document-heavy work with customizable model weights.

The model carries a 120B-A12B designation, with 120 billion total parameters and 12 billion active in each forward pass. A detailed count reaches 124.453 billion when it includes the vision encoder, projector and multi-token prediction module. The mixture-of-experts model has 512 routed experts per layer.

Its maximum context window is one million tokens. A hosted endpoint has a separate 262,000-token limit. The weights are available under the OpenMDW-1.1 license for deployment locally, in data centers or in the cloud. The supervised fine-tuning and reinforcement-learning data have not been published.

“Open models give enterprises a head start on building with their own data and intelligence,” said Joey Conway, NVIDIA’s senior director of Nemotron Models and Software.

Post-training results vary by task. In a 200-task indirect prompt-injection evaluation, a post-trained Nemotron model’s score rose from 35.19 to 76.16. It scored above GPT-6.1 Sol on that test.

A separate synthetic hospital-appeals evaluation showed different results across two setups. In the initial 16-case test, post-trained R1-super scored 31.3% on single attempts, compared with 35.2% for GPT-6 Sol and 16.4% for Claude Opus 5. In an expanded, exploratory 50-case comparison, R1-super scored 43.8%, while GPT-6 Sol scored 42.8%.

These evaluations measure customized systems on specific workflows. They do not establish that the unmodified Nemotron model broadly outperforms GPT or Claude models.