What happened
NVIDIA released Nemotron 3.5 Super VL, a 120-billion-parameter mixture-of-experts model that activates 12 billion parameters per token. The company says it handles text, images, documents and video alongside reasoning, instruction following and tool use, and can be post-trained with organizational data, policies, tools and workflows.
NVIDIA describes Nemotron 3.5 Super VL as an open, customizable for specialized enterprise agents. Its mixture-of-experts architecture has 120 billion total parameters, with 12 billion activated per token. The model is designed to perform reasoning, instruction following and tool use while understanding text, images, documents and video. NVIDIA positions it between smaller models for high-volume text work and larger frontier models for complex planning and orchestration.
The release includes base, post-trained and NVFP4-quantized checkpoints, training recipes, datasets and reinforcement-learning environments. Developers can use NVIDIA NeMo for supervised fine-tuning, low-rank adaptation and with their own data, tools, policies and workflows. NVIDIA is also releasing datasets through Hugging Face, including tool-use, finance, instruction-following, long-context, fairness and vision-language data. A separate dataset developed with IIT Madras and AI4Bharat contains about 2.8 trillion tokens across 22 Indian languages.
NVIDIA reports several performance results. It claims up to five times higher throughput than other open models in its size category for agentic tasks, an 80.8% score on the CharXiv chart-reasoning benchmark, or 84% with Python code interpretation, and up to 2.3 times faster token generation on DGX Spark using its LiLiCorr speculative-decoding method. The source also cites an independent EXO Labs evaluation placing the model on a Pareto frontier for accuracy and task-completion time. These figures are reported by NVIDIA, and the source does not include the underlying test conditions in full.
Source details: developer.nvidia.com ↗
Why it matters
The release targets a practical bottleneck in enterprise AI: agents that must repeatedly interpret documents and visual information, follow company rules, call tools and maintain context over multistep workflows. NVIDIA claims the model can combine these capabilities in one open-weight system, reducing the need to pass work between separate language and vision models. It also provides customization libraries, datasets and reinforcement-learning environments, which could make domain-specific deployment more accessible to organizations with the necessary infrastructure. Those benefits remain claims from NVIDIA or evaluations cited by the company; the source does not provide independent comparative testing across real enterprise deployments, pricing, or evidence that the model is broadly available without substantial engineering work.
Enterprise agents often need to combine structured and unstructured evidence, interpret charts or page layouts, preserve long conversational and tool context, and comply with organization-specific procedures. A single model covering those functions could simplify an architecture that otherwise uses separate text, vision and document systems. That is a practical product claim, not proof that one model will be more reliable or cheaper in every deployment.
The open-weight positioning gives organizations more control over local or private-cloud deployment and allows post-training on proprietary workflows. The source says supported environments include local infrastructure, data centers and cloud deployments, with NVIDIA NIM, Dynamo and AI Enterprise available for packaged serving and support. It does not specify licensing terms, total hardware requirements, hosted pricing or the amount of data and engineering needed to customize the model.
NVIDIA also presents the release as part of a broader model and tooling stack, with coalition contributions from Perplexity and Prime Intellect and evaluations across several agent harnesses. Those integrations may be relevant to developers building long-running agents, but the source does not establish production adoption, customer outcomes or a safety record. The cited benchmark results should therefore be treated as release claims or attributed evaluations rather than independently established industry performance.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Developers will need to assess the model on their own workflows, especially reliability, tool-use safety, latency, operating cost and performance on proprietary documents. The source documents access through supported local, data-center or cloud environments, Hugging Face, ModelScope, OpenRouter and NVIDIA deployment products, but it does not state a hosted price or establish general availability for every platform. NVIDIA’s benchmark claims, including throughput and chart-reasoning results, should be checked against reproducible methods and independent evaluations.
The immediate practical question is whether Nemotron 3.5 Super VL can maintain accuracy and policy compliance across long-running tool calls, not only answer individual multimodal questions. Organizations should test document extraction, chart interpretation, tool selection, recovery from errors and behavior under ambiguous or adversarial instructions.
Infrastructure requirements will shape who can use the model. NVIDIA says it is optimized for DGX Spark and provides deployment paths through NIM, Dynamo, Hugging Face, ModelScope and OpenRouter, but the source does not state a price, service-level commitment, access restrictions or whether each checkpoint is available through every route.
Independent replication is also important. The source supplies NVIDIA’s throughput, benchmark and deployment figures and cites EXO Labs for one evaluation, but it does not provide comparative results from independent enterprise users. Unknowns include real-world error rates, data governance implications, energy use, latency at different context lengths and performance across the model’s claimed language and industry domains.