NVIDIA opens a 120B multimodal model for agents: up to 5x throughput on one GPU, post-trainable on your own data

Enterprise agents no longer need separate text and vision-language models, nor task context passed between them. Handling text and visual tasks previously meant maintaining two models and moving task context between them. The model combines reasoning, instruction following, and tool use with text, image, document, and video understanding in one model, which NVIDIA says lets teams avoid deploying separate text and vision-language models.

A mixture-of-experts architecture keeps inference cost down by activating only part of the parameters per token. Total parameter count sets capacity, but activating all of it raises the cost of each inference step. Nemotron 3.5 Super VL has 120B total parameters and activates 12B per token, and NVIDIA reports up to 5x higher throughput on agentic tasks than other open models in its size class on a single GPU.

Organizations can post-train the model on their own data, tools, policies, and workflows rather than only calling general capabilities. A general model does not know an organization's terminology, tool interfaces, operating procedures, or policy requirements. Developers can use NVIDIA NeMo AutoModel for supervised fine-tuning and low-rank adaptation and reinforcement learning in agentic environments that reflect organizational workflows; NVIDIA also released agentic tool-use and conversational data, finance reasoning and tool-calling datasets, instruction-following and chat data, long-context transduction and translation data, fairness data, and 40 vision-language datasets.

On chart reasoning it scores above the smaller Nemotron 3 Nano Omni in the same family, and on Hermes Bench and LangChain behavioral evaluations it scores above the larger Nemotron 3 Ultra. Larger models are usually expected to be stronger on complex tasks, while the chart-reasoning comparison is against the smaller Nano Omni. It scored 80.8% on the CharXiv chart-reasoning benchmark versus 63% for Nemotron 3 Nano Omni, reaching 84% with Python code interpretation; it scored 63.44% on Hermes Bench and 87.05% on LangChain behavioral evaluations, above Nemotron 3 Ultra's 56.33% and 84.3%.

The report covers NVIDIA's release of Nemotron 3.5 Super VL with its post-training tooling and datasets, aimed at enterprise agent developers who need multimodal reasoning and tool use. The throughput and accuracy figures come from NVIDIA's own evaluations and apply to the stated single-GPU and deployment settings.

The up-to-5x throughput figure comes from NVIDIA's own evaluation. For Scale AI's vendor-credit processing workload at Skydance, inference cost will decrease by 40x and latency by 5x; Ramp expects up to 50% higher throughput and 10x lower latency for its receipt-to-structured-JSON workload; the 41% accuracy gain comes from AI21 and Alice's post-training result, with no uniform third-party reproduction conditions given. Comparisons such as CharXiv and Hermes Bench also come from NVIDIA's evaluation setups.