NVIDIA launches Nemotron 3.5 Super VL for multimodal agents

NVIDIA on Oct. 15, 2026, introduced Nemotron 3.5 Super VL, an open, customizable multimodal model designed for specialized enterprise agents that must read documents, inspect images and video, use tools and follow policy across longer workflows.

Evidence and context

According to NVIDIA’s technical blog published the same day, the model uses a 120 billion-parameter mixture-of-experts architecture with 12 billion active parameters per token. NVIDIA said the design is intended to balance model capacity with inference efficiency, and reported that on a single GPU the model delivers up to 5x higher throughput for agentic tasks than other open models in its size category.

The company said Nemotron 3.5 Super VL combines reasoning, instruction following and tool use with text, image, document and video understanding, allowing teams to handle multimodal tasks in one workflow rather than chaining separate text and vision models. NVIDIA also said developers can post-train the model with NeMo using proprietary data, tools, policies and workflows.

NVIDIA cited several customer-style deployments to illustrate the model’s intended use cases. Scale AI post-trained it for automated vendor credit processing at Skydance Corporation, which NVIDIA said will reduce inference cost for that workload by 40x and latency by 5x. Ramp is using it to convert receipt images into structured JSON, with NVIDIA saying Ramp expects up to 50% higher throughput and 10x lower latency. AI21 post-trained the model for Alice’s threat intelligence and vulnerability analysis, which NVIDIA said improved accuracy by 40%.

The company said the release includes base, post-trained and NVFP4-quantized checkpoints, along with training recipes, data and reinforcement-learning environments. NVIDIA also said it is releasing open datasets on Hugging Face covering agentic tool use, finance reasoning and tool-calling, instruction-following, long-context transduction, fairness and 45 vision-language datasets.

On benchmark comparisons provided by NVIDIA, Nemotron 3.5 Super VL scored 80.8% on CharXiv, versus 63% for Nemotron 3 Nano Omni, and 84% with Python code interpretation. NVIDIA also said the model scored higher than Nemotron 3 Ultra on Hermes Bench and LangChain behavioral evaluations, citing scores of 63.44% versus 56.33% on Hermes Bench and 87.05% versus 84.3% on LangChain.

For inference, NVIDIA highlighted LiLiCorr speculative decoding, NVFP4 quantization, multi-token prediction and context reuse across agent turns. The company said LiLiCorr can deliver up to 2.3x faster token generation than decoding without speculation on DGX Spark, and that NVIDIA Dynamo can serve up to 8x as many users while reducing first-token latency by 85% in a specific configuration on 16 NVIDIA B200 GPUs. Those figures were reported by NVIDIA and are configuration-specific rather than general guarantees.

The blog also said the model was distilled from Nemotron 3 Ultra and further trained with specialized text and vision teachers, with additional continued pretraining using Indic-language data. NVIDIA said it is releasing a dataset of about 2.8 trillion tokens across 22 Indian languages, developed with IIT Madras and its AI4Bharat lab.

OpenHands also published related material on Oct. 15, 2026, describing Nemotron 3.5 Super VL in visual agent workflows, while Multiverse Computing separately announced a multimodal model collaboration with NVIDIA the same day. NVIDIA’s post itself said the model is available through supported local, data center and cloud environments, including NIM and Dynamo deployment paths.