vLLM BlogSource publication:
Loading content…
NVIDIA released Nemotron 3.5 Super VL, an open multimodal model for long-running agents that folds document, image, and video understanding together with reasoning, instruction following, and tool use in one model, reducing the need to deploy separate text and vision models. It uses a 120-billion-parameter mixture-of-experts design with 12 billion active parameters, combined with a hybrid Mamba-Transformer architecture, multi-token prediction, and NVFP4 quantization, and can be post-trained with NVIDIA NeMo on an organization's own data, tools, policies, and workflows. The stated comparison is up to 5x faster throughput on agentic tasks than other open models in its size category on a single GPU, and 4x faster task completion.
Always-on agents can now interpret images, documents, and video and call tools with one model, instead of maintaining separate text and vision models. The prior common setup deployed a text model and a vision model separately, splitting multimodal understanding from high-accuracy execution. NVIDIA positions the model as the default high-accuracy agentic execution model within a system of models, alongside Nemotron 3.5 Lightning for high-volume text execution and Nemotron 3 Ultra for the hardest text-only reasoning.
Multimodal understanding and agentic execution sit in the same inference path, preserving task-relevant context across workflow steps. A hybrid Mamba-Transformer architecture, multi-token prediction, and NVFP4 quantization are used to improve generation efficiency and leave headroom for more concurrent long-context sessions. Stated deployment targets are 1x GB10, B200, B300, RTX PRO 6000, H200, and 2x H100, with production serving through NVIDIA NIM across cloud and data center environments.
During decoding, built-in multi-token prediction heads propose future tokens that the target model verifies, cutting the number of sequential generation steps for longer responses. NVIDIA states this accelerates token generation while preserving the target model's output quality. The stated comparison is up to 5x faster throughput on agentic tasks than other open models in its size category on a single GPU and 4x faster task completion, with over 600 TPS on Artificial Analysis Intelligence Index agentic tasks on a Baseten endpoint.
The model targets agent workflows that run continuously, call tools repeatedly, and handle documents, images, and video; NVIDIA lists customer service, document intelligence, enterprise productivity, video understanding, and web research as examples. Deployment targets include 1x GB10, B200, B300, RTX PRO 6000, H200, and 2x H100, with production serving through NVIDIA NIM and vLLM providing continuous batching, prefix caching, speculative decoding, and an OpenAI-compatible API. NVIDIA advises validating the NVFP4 launch configuration on the target hardware before use.
The throughput and task-completion multiples come from NVIDIA's own figures, measured on a single GPU and a Baseten endpoint, so values may differ on other hardware or serving stacks. A separate DGX Spark efficiency-frontier chart covers the related Nemotron 3.5 Super model and does not measure Nemotron 3.5 Super VL. The NVFP4 launch configuration is described as an example and needs validation on the target hardware.
