NVIDIA Nemotron 3.5 Super VL is now available on Baseten for inference and customization, the companies’ October 15 announcement says. The open 120B-A12B model combines visual understanding, reasoning and tool use for agents working with documents, images, application screens and video. Developers can evaluate it through Baseten’s Dedicated and Model APIs, then customize it with Baseten Loops.

The model uses a 120-billion-parameter Mixture-of-Experts design, activating 12 billion parameters per token. It builds on the Nemotron 3 Super architecture and combines Mamba and Transformer components for long-context processing; multi-token prediction is intended to speed generation. NVIDIA says NVFP4 KV cache quantization reduces memory needs for long-context inference. A vision encoder lets agents interpret visual material, including document layouts and screenshots, alongside text.

In NVIDIA’s reported text-benchmark results, Nemotron 3.5 Super VL scored 1,190 Elo on GDPval AA V2.1, compared with 493 for Nemotron 3 Super. It also recorded 47.3% accuracy on SciCode versus 40.8%, and 90.8% on GPQA-diamond versus 79.4%. The benchmarks cover professional tasks, scientific coding and graduate-level science questions, respectively.

Baseten Loops added training support on the model’s launch day. Loops is Baseten’s managed service for supervised fine-tuning and reinforcement learning; training runs use a Python script and the Loops SDK, while Baseten operates the distributed training and GPU infrastructure. Teams already training Nemotron 3 Super or Nemotron 3 Ultra on Loops can switch base models without changing tools, then deploy a trained checkpoint on Baseten. The company says the model is suited to customer service, document intelligence, enterprise productivity and web research.

NVIDIA Nemotron 3.5 Super VL is now available on Baseten for inference and customization, the companies’ October 15 announcement says. The open 120B-A12B model combines visual understanding, reasoning and tool use for agents working with documents, images, application screens and video. Developers can evaluate it through Baseten’s Dedicated and Model APIs, then customize it with Baseten Loops.

The model uses a 120-billion-parameter Mixture-of-Experts design, activating 12 billion parameters per token. It builds on the Nemotron 3 Super architecture and combines Mamba and Transformer components for long-context processing; multi-token prediction is intended to speed generation. NVIDIA says NVFP4 KV cache quantization reduces memory needs for long-context inference. A vision encoder lets agents interpret visual material, including document layouts and screenshots, alongside text.

In NVIDIA’s reported text-benchmark results, Nemotron 3.5 Super VL scored 1,190 Elo on GDPval AA V2.1, compared with 493 for Nemotron 3 Super. It also recorded 47.3% accuracy on SciCode versus 40.8%, and 90.8% on GPQA-diamond versus 79.4%. The benchmarks cover professional tasks, scientific coding and graduate-level science questions, respectively.

Baseten Loops added training support on the model’s launch day. Loops is Baseten’s managed service for supervised fine-tuning and reinforcement learning; training runs use a Python script and the Loops SDK, while Baseten operates the distributed training and GPU infrastructure. Teams already training Nemotron 3 Super or Nemotron 3 Ultra on Loops can switch base models without changing tools, then deploy a trained checkpoint on Baseten. The company says the model is suited to customer service, document intelligence, enterprise productivity and web research.