Nemotron 3.5 Super VL: NVIDIA Announces a 120B-Parameter Multimodal Model

A multimodal model for specialized agents

NVIDIA announced Nemotron 3.5 Super VL as an open, customizable model for workflows that combine reasoning, visual understanding, and tool use. According to the company’s technical description, it can process text, images, documents, and video within a single model, with the aim of avoiding the need to split some tasks between a language model and a separate vision model.

The model is intended for agents that must complete sequences of work, rather than simply answer a single question—for example, reading a document, cross-checking information, and using tools. NVIDIA also says organizations can adapt it with their own data, tools, and policies. These are stated use cases and capabilities; they do not prove that the system will perform equally well across industries or configurations.

The available evidence varies in scope. NVIDIA’s announcement confirms the release and places the model within the Nemotron family. The technical article provides figures and evaluation details, but it is material published by NVIDIA, not an independent replication. Its results should therefore be read as measurements reported by the manufacturer, not as a general performance guarantee.

MoE architecture: total capacity versus active computation

Nemotron 3.5 Super VL has 120 billion parameters in total and activates 12 billion per token, according to NVIDIA’s technical documentation. In a mixture-of-experts, or MoE, architecture, different specialized parts of the model can participate in processing an input. The active-parameter figure indicates how much of the model is used at each step; it does not mean that the system stores only 12 billion parameters.

That distinction can help balance capacity and computation: the model retains a large pool of parameters while activating a fraction of them for each token. But the active-parameter count alone does not determine actual memory requirements or speed. Numerical precision, context length, implementation, batch size, and hardware also matter.

The technical material also describes several inference techniques: LiLiCorr speculative decoding, multi-token prediction, context reuse, and quantization. These are different optimizations, and their effects depend on the workload. Any improvement reported for the full system should not automatically be attributed to the MoE architecture alone.

Benchmarks: useful results, tied to their test conditions

NVIDIA says the model achieves up to five times higher throughput on agentic tasks than other open models in its size category. The technical article also describes results on agent evaluations, chart understanding, and local execution on DGX Spark. These metrics are not equivalent: throughput measures token generation under particular conditions, while task completion also depends on the number of steps, tool calls, and tokens used.

For Hermes Bench, the technical article reports 63.44% for Nemotron 3.5 Super VL and 56.33% for Nemotron 3 Ultra. On LangChain behavioral evaluations, it reports 87.05% and 84.3%, respectively. Those figures are attributed to those specific tests; the source does not establish that they are independent evaluations or that they generalize to all agents, models, or datasets.

For chart understanding, NVIDIA reports 80.8% on CharXiv and 84% when Python code interpretation is used. The difference illustrates how results can depend on the evaluation procedure. To determine whether the model suits a real use case, it is worth reproducing the test with the organization’s own documents, tools, success criteria, and time limits.

DGX Spark, NVFP4, and costs: what cannot yet be compared

The technical documentation says Nemotron 3.5 Super VL was optimized for local execution on NVIDIA DGX Spark and cites an independent EXO Labs evaluation of the balance between accuracy and task completion time. However, the information provided does not include a complete, comparable table of latency and throughput against Bolt 74B-A9B or Bolt E4B, or consistent power-consumption measurements for DGX Spark, RTX 6000, and RTX 8000.

It is therefore not possible to rigorously determine the cost of each task on those GPUs or which model would be faster in a particular deployment. Results depend on the exact hardware, weight format, inference engine, input size, concurrency, and definition of task completion. The announcement of the Bolt models does not, by itself, amount to a performance comparison with Super VL.

NVIDIA describes an NVFP4 recipe for the Blackwell architecture that quantizes weights and activations to four bits, and says quantization-aware distillation is intended to preserve accuracy compared with the BF16 model. It also says the same checkpoint can run on non-Blackwell platforms with four-bit weights and 16-bit activations. This describes compatibility and methodology; the available sources do not provide an independent measurement on a common workload quantifying the resulting change in accuracy or energy use.

Compatibility and potential industrial uses

NVIDIA presents its own tools for customizing and deploying the model, including NeMo, NVIDIA NIM, and Dynamo. The technical documentation also lists a broad set of ecosystem platforms and frameworks, including vLLM, SGLang, Ollama, and LM Studio. A tool’s presence in a partner or framework list does not prove that every feature, format, and optimization is available in every version; deployment teams should check support for the specific checkpoint and backend.

Potential uses described include document intelligence, video analysis, web research, cybersecurity, and programming assistance. These are possible applications, not guaranteed outcomes. Each use case also requires evaluation of sensitive-data handling, response reliability, access policies for tools, and recovery from errors.

The documentation mentions adaptations for specialized workflows, including finance and threat analysis. The specific improvements reported apply to particular configurations and post-training efforts; they should not be treated as automatic properties of the base model. Industrial adoption should be validated using organization-specific data and measures of security, quality, and cost.

Conclusion: promising capabilities, but limited evidence

Nemotron 3.5 Super VL combines 120 billion total parameters with 12 billion active per token and a multimodal design aimed at agentic workloads. NVIDIA’s documentation describes techniques for accelerating inference and reports results on specific benchmarks; it also outlines local execution on DGX Spark and options for customization.

A cautious conclusion is that the model is worth evaluating for tasks that combine text, documents, or video with reasoning and tool use. But the evidence provided does not support claims that it is generally the fastest, consumes less energy than specific alternatives, or has a known cost on RTX 6000, RTX 8000, or DGX Spark. Nor are specific, consistent comparisons against Bolt 74B-A9B and Bolt E4B provided.

For readers comparing systems, the practical choice is to measure the complete task, not just tokens per second: accuracy, end-to-end latency, power use, cost, stability, and actual compatibility. The architecture and announced figures can help formulate hypotheses; testing on your own data and hardware shows whether they hold.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.